{
  "id": "data-visualization",
  "title": "Data Visualization",
  "lang": "en",
  "source_lang": "en",
  "summary": "Data visualization represents data graphically to support exploration, analysis, explanation, and communication.",
  "category": "computer-science",
  "subcategory": "software",
  "type": "concept",
  "aliases": [
    "Data visualisation"
  ],
  "generated_at": "2026-10-25",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/data-visualization/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/data-visualization/",
      "json": "https://aiwiki.page/zh/data-visualization.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Data visualization is the representation of data through graphical forms, including charts, maps, diagrams, and interactive displays. It maps values, categories, and relationships to visual properties such as position, length, size, and color, allowing people to examine patterns that may be difficult to recognize in tables or text. Closely connected to statistics, computer science, and visual design, it supports both exploratory analysis and the communication of findings. A visualization can reveal distributions, trends, associations, and anomalies, but does not by itself establish causation or guarantee that the underlying data are accurate.\n\nThe field draws on traditions of statistical graphics, cartography, and scientific illustration, and has expanded through computing into interactive visualization and visual analytics. Its methods range from familiar bar charts and time-series plots to geographic displays, network diagrams, and three-dimensional representations of scientific data. Effective visualization depends on the characteristics of the data, the questions being asked, human perception, and the context of use. Its practice therefore encompasses data preparation, visual encoding, interaction, software implementation, evaluation, accessibility, and the responsible presentation of uncertainty and evidence."
    },
    {
      "heading": "Scope and Relationships to Other Fields",
      "text": "Data visualization is the systematic mapping of data attributes and relationships to visual representations that support inspection, reasoning, or communication. Three elements are distinct: the data, such as monthly rainfall measurements; the representation, such as a line whose horizontal position encodes month and vertical position encodes rainfall; and interpretation, the activity through which a viewer identifies patterns or answers questions. The mapping establishes which visual differences correspond to differences in the data. Changing an axis scale or aggregating daily measurements into monthly totals changes that correspondence, even when the underlying records remain unchanged. A visible pattern is therefore a property of data as represented, not an unmediated view of the phenomenon being measured.\n\nThe field encompasses several overlapping traditions. Statistical graphics, closely associated with statistics, emphasizes distributions, relationships among variables, and the examination of statistical models. A scatterplot of measured height against weight belongs to this tradition. Information visualization commonly addresses abstract data without an inherent spatial arrangement, including organizational hierarchies, citation networks, and tables of business transactions. Its designers must establish meaningful spatial arrangements rather than simply reproduce physical coordinates. Scientific visualization often concerns measurements or simulations situated in physical space, such as temperature throughout an atmospheric model or velocity in a fluid-flow simulation. These distinctions describe customary emphases rather than exclusive categories: an environmental analysis may combine a geographic field, statistical distributions, and a network of monitoring stations within the same interface.\n\nVisualization also intersects with illustration, art, and graphic design, but these activities are not interchangeable. An illustration may depict a machine’s construction without systematically encoding observations; decorative imagery may establish mood without representing data at all. Conversely, a diagram can qualify as data visualization when its marks encode recorded entities and relations, even if it contains no numerical axis. The term infographic usually refers to a broader communicative composition that may combine charts, maps, explanatory text, icons, photographs, and illustrative diagrams. Only some of these components necessarily encode data. The relevant distinction is not whether an image is attractive or pictorial, but whether identifiable rules connect its visible features to the information represented.\n\nVisual analytics integrates interactive visual interfaces with computational analysis. James J. Thomas and Kristin A. Cook’s 2005 research agenda, Illuminating the Path, characterized it as analytical reasoning facilitated by interactive visual interfaces. In a typical workflow, an algorithm might group thousands of documents by textual similarity; an analyst then inspects the groups, filters by date, examines individual records, and revises the analysis. Computational methods provide operations such as search, aggregation, and clustering, while visualization makes their outputs available for scrutiny and human-directed exploration. An interactive chart alone does not necessarily constitute visual analytics: the term emphasizes an integrated analytical process, rather than interaction as an isolated interface feature.\n\nGraphics, tables, and numerical summaries serve complementary purposes. Tables support exact lookup and preserve explicit values; summaries compress observations into quantities such as means and standard deviations; graphics expose configurations that those quantities may conceal. Francis Anscombe demonstrated this distinction in 1973 with four datasets, each containing 11 paired observations and having nearly identical means, variances, correlations, and fitted regression equations. Their scatterplots nevertheless show markedly different structures, including curvature and influential observations. The example does not establish that graphics replace numerical analysis. It demonstrates why identical summary statistics need not imply equivalent data structure, and why checking a visual pattern against values, definitions, and statistical calculations remains part of interpretation."
    },
    {
      "heading": "Historical Development",
      "text": "Antecedents of data visualization include cartography, astronomical diagrams, surveying, and graphical records of movement and time. These traditions established ways to relate positions on a surface to locations, sequences, or measured quantities, although they did not constitute a single discipline. In the fourteenth century, Nicole Oresme used geometric diagrams in which the length of a figure represented an extension such as time and its height represented the intensity of a quality, including velocity. René Descartes’s La Géométrie (1637) helped establish algebraic treatments of geometry that subsequently supported coordinate-based graphics. Joseph Priestley’s A Chart of Biography (1765) placed lifespans along a common chronological scale, making contemporaneity and historical duration visually comparable.\n\nWilliam Playfair made quantitative comparison central to commercially published statistical graphics. His Commercial and Political Atlas (1786) used line charts to show economic quantities over time and included a bar chart comparing Scotland’s imports and exports with different trading partners. Unlike a geographical map, this bar chart assigned positions to categories rather than physical locations; bar lengths represented monetary amounts. In The Statistical Breviary (1801), Playfair introduced circular statistical displays, including the chart commonly identified as the first pie chart. His graphics joined numerical scales, labels, and explanatory text to arguments about trade and national resources, demonstrating that economic relationships could be communicated through designed visual comparisons rather than tables alone.\n\nNineteenth-century public-health graphics connected geographical patterns and mortality statistics to investigations and administrative reform. John Snow’s map of deaths associated with the 1854 cholera outbreak in London’s Soho district represented deaths with short bars at addresses and marked public water pumps. Published in his expanded On the Mode of Communication of Cholera (1855), it supported his account of transmission through contaminated water, particularly the Broad Street pump. The familiar map was not, by itself, the evidence that caused the pump handle’s removal on 8 September 1854: Snow also investigated households, water consumption, and apparent exceptions, and the outbreak was already declining. Florence Nightingale’s mortality diagrams, published in 1858, used polar-area displays to compare monthly deaths among British soldiers during the Crimean War. Sector areas, rather than radii, encoded deaths; colors distinguished causes. The diagrams supported her case for sanitary reform, alongside statistical tables, reports, and institutional advocacy.\n\nCharles Joseph Minard’s 1869 depiction of Napoleon’s Russian campaign combined several quantitative relationships in one composition. A band traced the army’s route during the 1812–1813 campaign, with width proportional to troop strength and contrasting colors distinguishing advance and retreat. Place names and geographical positions supplied context, while a temperature plot beneath the retreat route recorded temperatures on the Réaumur scale and associated dates. The main band narrowed from 422,000 men at the campaign’s opening to 10,000 at its return crossing of the Niemen. These are the figures represented by Minard, not a complete accounting of every military movement or cause of loss. The graphic juxtaposed attrition, movement, and weather without independently establishing their causal relationships.\n\nTwentieth-century work increasingly addressed visualization as a systematic method of representation and analysis. Jacques Bertin’s Sémiologie graphique (1967) examined how planar position and visual variables—including size, value, texture, color, orientation, and shape—could represent different kinds of information. John Tukey’s Exploratory Data Analysis (1977) consolidated an approach developed through earlier work: graphical examination should help identify patterns, unusual observations, and possible models before analysis is restricted to confirmatory procedures. Stem-and-leaf displays and box plots exemplified this emphasis on exposing distributional structure.\n\nDigital computer graphics changed both production and use. Ivan Sutherland’s Sketchpad (1963) demonstrated interactive manipulation of graphical objects, while later statistical systems enabled brushing, linked views, and dynamic exploration. The 1987 report Visualization in Scientific Computing helped consolidate scientific visualization as a research agenda. From the 1990s, the World Wide Web expanded distribution beyond installed applications; subsequent browser technologies supported interactive graphics within documents. D3, introduced in 2011, made data-driven manipulation of web-document elements a widely used approach, connecting visualization to JavaScript and established web standards."
    },
    {
      "heading": "Data Types, Structure, and Preparation",
      "text": "An attribute’s type determines which distinctions and operations its values support. Categorical attributes identify groups; nominal categories, such as country names, have no intrinsic order, whereas ordinal categories, such as low, medium, and high, have an order without necessarily equal spacing. Quantitative attributes represent numerical magnitudes, but their interpretation depends on the measurement scale. In S. S. Stevens’s 1946 classification, interval scales have meaningful differences but an arbitrary zero: 20 °C is 10 degrees warmer than 10 °C, not twice as hot. Ratio scales, such as duration and mass, also support meaningful ratios. Numeric storage does not establish quantitative meaning: postal codes remain identifiers. Temporal attributes include instants, durations, and recurring calendar categories; months have unequal lengths, and local timestamps can repeat when clocks move backward. Spatial attributes describe locations, extents, or spatial relationships; coordinates require a reference system, since degrees of latitude and longitude are not uniform units of ground distance.\n\nData structure is distinct from attribute type. A table organizes records into rows and attributes into columns, with keys identifying records and connecting related tables in a database. Multidimensional arrays index values along several dimensions: a temperature array indexed by time, altitude, latitude, and longitude requires selecting, slicing, or reducing dimensions for many displays. Hierarchies encode parent–child relationships, while networks allow more general connections represented by graphs. A strict tree gives each non-root node one parent; a transport network can contain cycles and multiple routes. Fields assign values to positions in a domain, as in a temperature field sampled on a grid. Geometric data explicitly represent objects such as points, lines, surfaces, and volumes. These structures constrain displays: connecting unordered table rows with lines introduces a sequence absent from the data, while drawing a field as a continuous surface introduces assumptions about values between samples.\n\nPreparation changes the represented data, not merely their formatting. Cleaning can standardize category spellings, parse dates, resolve duplicates, and distinguish invalid values from legitimate extremes. Joins combine records through matching keys, but their cardinality matters. If one table contains two records for a location and another contains three, joining on location alone produces six matched rows, potentially multiplying totals. Reshaping from wide form—one column per year—to long form—year and value columns—makes time an explicit attribute without necessarily changing observations. Aggregation does change observational detail: hourly readings reduced to daily means no longer expose within-day peaks. Unit reconciliation is equally substantive. Converting Fahrenheit to Celsius requires \\(C=(F-32)\\times5/9\\), not simple rescaling, and monetary series may require distinguishing nominal amounts from inflation-adjusted values.\n\nThe observational unit specifies what one record represents: a person, household, transaction, sensor reading, or regional annual total. A table containing several visits per person cannot yield a person count by counting rows. Denominators likewise define the quantity displayed. In a hypothetical comparison, 50 events among 1,000 observations equal 5%, whereas 100 among 10,000 equal 1%; counts and proportions therefore produce opposite rankings. Combining subgroup percentages requires their denominators: the pooled proportion is \\(\\sum_i x_i/\\sum_i n_i\\), not generally the unweighted mean of \\(x_i/n_i\\). Metadata document variable definitions, units, category codes, geographic reference systems, collection periods, and coverage. Provenance records sources and processing steps, including filters, joins, and aggregation rules, supporting reproducibility and explaining differences between derived datasets.\n\nRecorded zero, missingness, and a value below a detection threshold represent different states. Zero records an observed absence or measured zero under the variable’s definition; missing data record no usable value. A concentration reported as below 0.1 mg/L is left-censored: its exact magnitude is unresolved, rather than known to be zero. Replacing all three states with zero changes distributions and averages. Omitting missing records changes denominators, while connecting a time series across missing intervals can imply uninterrupted observation. Imputation supplies estimated values rather than recovered measurements; their status and the method used belong in the metadata and provenance."
    },
    {
      "heading": "Analytical Tasks and the Design Process",
      "text": "Analytical tasks describe what a reader or analyst needs to accomplish, rather than which chart should be drawn. Lookup retrieves a particular value; comparison examines differences between observations or groups; and ranking orders items by an attribute. Distribution assessment concerns spread, concentration, skewness, or multiple modes. Relationship detection asks how attributes vary together, while anomaly identification seeks observations that depart from an expected pattern. Examination of change concerns differences across time, including trends, abrupt shifts, and recurring cycles. These tasks often form sequences: an analyst might identify an unusual month, compare it with earlier months, and then inspect the contributing categories. Robert Amar, James Eagan, and John Stasko’s 2005 taxonomy formalized ten low-level analytical tasks, including retrieving values, filtering, finding extrema, sorting, characterizing distributions, and finding anomalies.\n\nTask abstraction translates a domain question into operations on data and comparisons that a display must support. For example, “Which factories reduced electricity consumption most between 2020 and 2024?” requires identifying factories as the entities, selecting two time points, calculating a change for each factory, and ranking the results. Absolute change, \\(E_{2024}-E_{2020}\\), answers a different question from proportional change, \\((E_{2024}-E_{2020})/E_{2020}\\); the latter is undefined when the baseline is zero. A comparison of energy efficiency would additionally require an appropriate production denominator. The abstraction therefore exposes decisions that wording alone can conceal. It also distinguishes the analytical purpose from the interaction used to achieve it: filtering may help locate a factory, but filtering is not itself the substantive question.\n\nTask descriptions must account for whether the target and its location are already known. Looking up a named factory differs from discovering an unexpectedly large reduction among thousands of factories. Matthew Brehmer and Tamara Munzner’s 2013 multi-level task typology separates why a task is performed, how it is performed, and what its inputs and outputs are. Its search categories distinguish lookup, locate, browse, and explore according to knowledge of the target and its location. This makes it possible to describe an analytical workflow without prematurely prescribing a chart or interface. Detecting a relationship, for example, may involve visual comparison and estimation of correlation, but does not by itself establish causation.\n\nDesign typically proceeds iteratively through investigation, requirements, alternatives, prototypes, and revision. Interviews and observation establish users’ responsibilities, existing practices, terminology, and operating constraints. Requirements then translate these findings into concrete capabilities: compare specified periods, retain selected entities across views, or retrieve exact values for a report. Sketches allow several arrangements to be considered before implementation; prototypes make assumptions about navigation, data availability, and response time testable. A factory analyst preparing a monthly report may require stable comparisons and exportable tables, whereas an operator investigating a current deviation may prioritize recent observations and rapid filtering. Revisiting requirements after prototype use can reveal that an apparent chart-selection problem is actually a mismatch in available data or task definition.\n\nTamara Munzner’s 2009 nested model of visualization design and validation distinguishes four levels: domain problem and data characterization; data and operation abstraction; visual encoding and interaction design; and algorithm design. The levels are dependent rather than interchangeable. A system can render quickly yet encode the wrong comparison; it can support an accurate comparison that does not address the user’s actual problem. Corresponding checks therefore ask different questions: whether the domain problem has been understood, whether abstractions preserve its relevant meaning, whether visual and interactive choices support the required tasks, and whether the implementing algorithms meet computational demands. Findings at an inner level can prompt revision at an outer one, rather than merely further optimization of the existing implementation."
    },
    {
      "heading": "Visual Perception and Cognition",
      "text": "Selective attention determines which parts of a display receive detailed processing. A red point among gray points can attract attention rapidly, whereas finding a particular conjunction, such as a red triangle among red circles and blue triangles, often requires more deliberate search. Anne Treisman and Garry Gelade’s 1980 feature-integration theory distinguished the registration of basic visual features from the attentional processes needed to bind features into objects. Features commonly described as preattentive include differences in color, orientation, size, and motion, but they do not constitute a context-independent checklist. Detection depends on contrast, distractor similarity, crowding, the observer’s goal, and competing features. A highlighted point may be conspicuous in a sparse scatterplot but difficult to locate in a dense, multicolored one; rapid detection also does not imply accurate numerical interpretation.\n\nGrouping principles associated with Gestalt psychology explain why viewers perceive structure beyond individual marks. Proximity encourages nearby points or labels to be read together; similarity groups marks sharing color or shape; enclosure makes items within a boundary appear to form a set. Connection can make linked points appear related even when other points are closer, while good continuity encourages viewers to follow smooth paths through intersections. These effects can reinforce or conflict with data relationships. For example, spacing rows into blocks suggests categories even without headings, and connecting observations suggests an ordered sequence rather than an unordered collection. In a line chart, viewers may follow a continuous trajectory more readily than identify which segments share a particular color. Perceived groups therefore depend partly on graphical organization, not solely on values.\n\nWilliam S. Cleveland and Robert McGill’s 1984 study of graphical perception investigated elementary judgments underlying statistical charts. Their experiments asked observers to estimate relative magnitudes encoded through graphical forms and compared the resulting errors. Position along a common scale generally supported more accurate judgments than encodings requiring less directly aligned comparisons; angle and area judgments were typically less accurate than the strongest positional encodings. This helps explain why two nearby dot positions can support a finer comparison than two similarly sized pie sectors. The research distinguished tasks such as judging position, length, angle, and area rather than treating a chart as an indivisible object. Its findings support empirical tendencies, not a universal ranking valid for every dataset, display size, or analytical task.\n\nWorking memory limits the amount of information that can be maintained while comparing different parts of a display. Nelson Cowan’s 2001 review argued for a capacity of roughly four chunks under conditions that restrict rehearsal and other memory supports; this is not a fixed limit of four chart marks. Expertise and meaningful grouping change what constitutes a chunk. Comparing values across separate screens requires retaining earlier values, their categories, and possibly their units. Keeping related panels simultaneously visible can externalize some of this information, although dense arrangements introduce additional visual-search demands. Search becomes harder when targets resemble distractors or when labels must repeatedly be matched to distant legends.\n\nChange blindness further limits comparisons across time. In experiments published in 1997, Ronald Rensink, J. Kevin O’Regan, and James Clark alternated scene images with brief blank intervals and showed that observers could miss substantial changes. Interrupting the visual transient that would normally draw attention makes change detection depend more heavily on directed search. In a dashboard, replacing one view with another may therefore leave an altered point or category unnoticed, particularly if its location was not already attended. Motion can attract attention, but simultaneous movements may compete rather than make every change evident.\n\nPrior knowledge affects both recognition and interpretation. Familiar chart conventions reduce decoding effort, while domain expertise supplies expectations about plausible ranges and relationships. Expectations can also contribute to cognitive biases: confirmation bias may direct attention toward observations supporting an existing explanation, and anchoring may influence judgments relative to an initially encountered value. These influences are distinct from sensory accuracy: a viewer can correctly read a position yet select unrepresentative evidence or interpret a visible association as causal."
    },
    {
      "heading": "Visual Encoding and Graphical Composition",
      "text": "A visual encoding maps data attributes to properties of graphical marks. Points identify locations or individual observations; lines connect positions or indicate boundaries; areas represent regions or quantities through their extent. Their visual channels include position, length, orientation, size, shape, luminance, and hue. These channels can operate together: a point may encode two measurements through horizontal and vertical position, a category through shape, and another quantity through area. The mappings are distinct from the marks themselves. A line’s position can represent a measured value, while its hue identifies a series and its thickness represents a separate quantity.\n\nExpressiveness concerns whether a representation preserves the structure of the data without introducing unsupported relationships. Distinct shapes can identify unordered categories, whereas a sequence from light to dark suggests an ordering. Assigning that sequence to unordered departments may imply a ranking that the data do not contain. Effectiveness concerns how well an expressive encoding supports a particular task. Comparing values against a common positional scale generally permits more precise judgments than comparing areas, although suitability depends on the task, display, and audience. Jock Mackinlay’s 1986 work on automated graphical presentation formalized expressiveness and effectiveness as criteria for selecting graphical designs. Encoding every available attribute is not necessarily effective: additional channels can compete for attention or make a legend difficult to decode.\n\nA scale translates data values into channel values. A linear positional scale has the form \\(p(x)=ax+b\\), so equal numerical differences produce equal distances. A logarithmic scale instead uses \\(p(x)=a\\log(x)+b\\): values of 1, 10, and 100 are equally spaced because their ratios are equal. Ordinary logarithmic scales cannot represent zero or negative values; square-root and symmetric logarithmic transformations offer different mappings, whose definitions must remain clear. Several design decisions are separate from the transformation. Axis limits determine the displayed range; tick spacing determines the frequency of reference marks; and a positional baseline supplies the origin from which an extent is read. Bars encode quantities through lengths from a baseline, normally zero, whereas point positions can be compared within a restricted range. Aspect ratio changes apparent slopes without changing observations. Cartesian and polar coordinate systems also produce different geometries: in polar coordinates, equal angular intervals occupy greater arc lengths farther from the center.\n\nColor schemes should match the attribute’s structure. Qualitative schemes distinguish categories without suggesting order. Sequential schemes represent increasing values through a perceptually ordered progression, often emphasizing changes in lightness rather than hue alone. Diverging schemes combine two progressions around a meaningful reference, such as zero change, a target, or a long-term average; their center need not be the arithmetic midpoint of the observed range. A rainbow progression can introduce apparent boundaries and uneven emphasis because equal numerical steps need not produce equal perceptual changes. Size mappings require similar care. For a circular symbol representing a positive quantity \\(v\\), proportional area requires \\(\\pi r^2=kv\\), hence \\(r=\\sqrt{kv/\\pi}\\). Doubling the value therefore multiplies the radius by \\(\\sqrt{2}\\), not by two; doubling the radius would quadruple the area.\n\nGraphical composition establishes how these encodings are read together. Legends explain mappings, while axis titles identify variables, units, and transformations. Direct labels can reduce repeated movement between a legend and the plotted marks, provided they do not obscure observations. Alignment supplies shared reference positions; whitespace separates groups without adding marks; and visual hierarchy distinguishes data, annotations, and supporting guides. Heavy gridlines can compete with the encoded values, while overly faint reference marks can make estimation difficult. Titles and annotations provide context but do not replace explicit units or scale definitions.\n\nSmall multiples repeat the same graphical design across subsets, such as one panel per region or year. Consistent scales allow differences in magnitude to be compared directly; independently scaled panels emphasize within-panel variation but can make unequal magnitudes appear similar. Panel order, dimensions, spacing, and repeated reference lines affect whether comparisons remain local or extend across the entire arrangement. Shared legends and selective repetition of axis labels can reduce clutter while preserving enough information to interpret each panel."
    },
    {
      "heading": "Charts for Comparisons, Distributions, and Time",
      "text": "Bar charts compare quantities across categories through lengths measured from a common baseline; dot plots encode the same quantities by position without filling the intervening space. Bars ordinarily begin at zero because their lengths represent magnitude, whereas dot plots can use a restricted range if clearly labeled. Sorting categories by value makes rank and extremes visible, but alphabetical or domain-specific order may better support lookup. Grouped bars place subcategories side by side, supporting comparisons both within and across groups. Stacked bars emphasize totals and composition: only the first segment consistently shares a baseline, making comparisons of intermediate segments harder. A 100% stacked chart compares proportions but hides differences in total size; identical halves could represent 5 of 10 observations or 500 of 1,000.\n\nHistograms divide a quantitative variable into intervals and represent the observations within each bin. Bin width and boundary placement can reveal or obscure gaps, skewness, and multiple peaks. With unequal bin widths, bar area—not height alone—must represent frequency or probability; a probability-density histogram uses height \\(n_j/(nw_j)\\), where \\(n_j\\) is the bin count, \\(n\\) the sample size, and \\(w_j\\) the bin width. Density plots replace bins with a smooth estimate, commonly using kernel density estimation. Their bandwidth controls smoothing: small values can produce spurious peaks, while large values can merge distinct groups. Smoothing can also extend beyond physically possible boundaries unless the estimator accounts for them.\n\nBox plots compactly summarize a distribution through its median, quartiles, and whiskers. In the convention popularized by John Tukey’s Exploratory Data Analysis (1977), the box spans the first and third quartiles; whiskers reach the most extreme observations within \\(1.5\\) interquartile ranges of the box, and more distant observations appear separately. These points are not automatically errors. Other conventions use minimum and maximum values or selected percentiles, so whisker definitions matter. Violin plots display mirrored density estimates, sometimes with an embedded box plot; their shapes depend on smoothing, and widths need not encode sample size. An empirical cumulative distribution function plots the fraction of observations at or below each value, allowing medians, percentiles, and threshold exceedances to be read without binning or bandwidth selection.\n\nScatterplots place paired quantitative measurements at coordinates \\((x_i,y_i)\\), preserving which values belong to the same observation. They expose nonlinear relationships, changing spread, clusters, and unusual points that a single correlation coefficient can conceal. Overlapping observations may require transparency, displacement, or aggregation, with displacement distinguished from measured variation. A regression line or smooth trend overlay summarizes a fitted relationship rather than replacing the observations. An isolated point can reflect error, a rare valid case, or an omitted grouping variable. Neither proximity to a fitted line nor a strong association establishes causation; shared influences, selection, and temporal ordering require separate analysis.\n\nLine charts conventionally place time horizontally and connect successive observations, supporting examination of trends, cycles, and changes. Straight segments imply interpolation between measured times, not continuous observation. Irregular sampling therefore requires spacing proportional to elapsed time; missing intervals should not silently become evidence of uninterrupted behavior. Step charts represent quantities held constant until an event, such as a posted price, with the step position indicating when the new value takes effect. Connected scatterplots instead use two measured variables as coordinates and join observations in temporal order; arrows or dates clarify direction. Horizon charts compress signed time series by dividing magnitude into bands, overlaying those bands, and encoding magnitude through color intensity, while distinguishing positive and negative values. This saves vertical space but makes exact readings dependent on the band scale and legend.\n\nPie charts encode parts of a total through sector angles and areas. Their denominator must be meaningful, with mutually exclusive categories covering the stated whole; overlapping survey responses cannot be treated as ordinary slices of one population. Similar sectors are difficult to compare precisely, especially across separate pies without a shared baseline. Bars support more direct comparisons, while stacked bars and rectangular part-to-whole displays emphasize composition. Omitting categories or renormalizing only the displayed subset changes the whole being represented and must be explicit."
    },
    {
      "heading": "Multivariate and High-Dimensional Visualization",
      "text": "A scatterplot matrix arranges pairwise plots of quantitative variables in a grid, exposing associations, outliers, and differences between groups. With \\(p\\) variables, there are \\(p(p-1)/2\\) distinct pairs; a full matrix usually repeats them across its diagonal and places variable names or univariate distributions on the diagonal. The classic iris dataset, analyzed by Ronald Fisher in 1936, contains 150 flowers, three species, and four measurements of sepal and petal dimensions. Its six distinct pairwise plots show that species separation depends on which measurements are compared. Such matrices do not directly reveal relationships that depend jointly on three or more variables.\n\nParallel coordinates represent each observation as a polyline crossing parallel axes, with its intersection on each axis encoding one attribute. Bundles can expose shared profiles, while crossings between adjacent axes can indicate reversals in relative ordering. Axis order determines which relationships are directly visible: a display with \\(p\\) axes exposes only \\(p-1\\) adjacent pairs at once. Axis scaling also matters, because normalization can emphasize relative profiles while obscuring absolute magnitudes. Heatmaps instead encode matrix values through color, often using rows for observations and columns for variables. Row-wise or column-wise standardization changes the comparison being made; a gene-expression heatmap standardized by gene emphasizes relative variation across samples rather than absolute expression between genes. Glyph-based displays assign several attributes to components of a repeated symbol—such as spoke lengths in star glyphs—but shape interactions and limited space complicate precise comparisons.\n\nFaceting and conditioning separate observations into panels according to categorical attributes or intervals of a quantitative attribute. For example, plotting income against age in separate panels by region permits comparison of within-region relationships without adding further visual channels to one crowded plot. Shared axes support magnitude comparisons, whereas independently scaled axes reveal local variation but weaken direct cross-panel comparison. Ordering rows, columns, axes, or panels can expose gradients and blocks; sorting a heatmap by hierarchical clustering produces a different view from chronological or alphabetical ordering. Coordinated panels retain complementary representations of the same records, allowing a subgroup identified in one view to be examined in others.\n\nDimensionality reduction replaces many attributes with a smaller set of coordinates. Principal component analysis (PCA) constructs orthogonal linear directions that successively maximize variance. Under the usual Euclidean formulation, projection onto the first components also minimizes squared reconstruction error among linear subspaces of the same dimension. A two-component plot therefore preserves selected large-variance patterns, not necessarily rare features or distinctions between known classes. Centering and scaling are consequential: applying PCA directly to variables measured in kilograms and millimeters can make unit-dependent variance dominate the result. Reporting explained-variance proportions identifies how much total variance the displayed components retain, but does not establish that they preserve every analytically important relationship.\n\nt-SNE, introduced by Laurens van der Maaten and Geoffrey Hinton in 2008, converts high-dimensional similarities into probabilities and minimizes their divergence from similarities in a low-dimensional layout. Its asymmetric objective strongly penalizes separating observations that were close in the original representation; perplexity controls an effective neighborhood scale. UMAP, described by Leland McInnes, John Healy, and James Melville in 2018, constructs a weighted neighborhood graph and optimizes a low-dimensional representation of that graph. Its neighborhood-size parameter affects the balance between local and broader structure, while minimum-distance settings influence the compactness of the embedding. Neither method guarantees faithful preservation of all original distances, densities, or global geometry.\n\nConsequently, an isolated island in a projection is not by itself evidence of a discrete population, and empty space between islands is not a calibrated measure of dissimilarity. In t-SNE, apparent cluster sizes and between-cluster distances can be particularly misleading. Both nonlinear methods can change with random initialization, distance metric, preprocessing, and parameter settings. Comparisons across runs, inspection of original-variable distributions, and checks of neighborhood preservation provide evidence about which patterns persist. Crowding remains fundamental: many high-dimensional neighborhood relationships cannot simultaneously fit into two dimensions without distortion, just as dense scatterplot matrices and parallel-coordinate displays eventually exceed available screen space."
    },
    {
      "heading": "Geographic and Spatiotemporal Visualization",
      "text": "Geographic displays associate observations with locations, routes, or regions on a map. Choropleth maps encode values through the fill color or shading of geographic units. They commonly represent rates, proportions, or densities: population density is \\(D_i=P_i/A_i\\), where \\(P_i\\) is population and \\(A_i\\) is area. Mapping raw population counts this way can conflate the number of people with the size of administrative units. Proportional-symbol maps instead scale symbol area to a quantity; for circles, radius must therefore vary with the square root of the value. Dot-density maps assign each dot a fixed quantity, such as 1,000 residents, but dots placed within census districts usually indicate an allocation within those districts, not individual addresses. Cartograms resize geographic regions according to population or another variable, sacrificing some combination of shape, distance, and adjacency to support quantitative comparisons.\n\nA map projection transforms locations on the curved Earth onto a plane; no projection preserves area, shape, distance, and direction everywhere. Equal-area projections preserve relative geographic areas, whereas the Mercator projection increasingly enlarges areas toward the poles. Consequently, a choropleth on Mercator can give high-latitude regions disproportionate visual prominence even when their numerical values are encoded correctly. Geographic extent also affects interpretation: changing the displayed region changes the apparent context, while automatically recalculating color classes can make identical values look different across panels. Large, sparsely populated districts occupy more screen space than small, densely populated ones, so the visual prominence of a color need not correspond to the share of people experiencing the mapped condition.\n\nSpatial aggregation changes both displayed patterns and the statistics calculated from them. The modifiable areal unit problem comprises a scale effect, arising when observations are aggregated into larger or smaller units, and a zoning effect, arising when boundaries change at a comparable level of aggregation. Stan Openshaw’s 1984 monograph helped establish this distinction in spatial analysis. A concentration visible in neighborhood data may disappear in county averages, while alternative district boundaries may change correlations between mapped variables. The ecological fallacy is a separate inferential error: relationships among areas do not necessarily describe relationships among individuals within them. William S. Robinson demonstrated this problem in 1950 using aggregate and individual relationships involving literacy and foreign birth in the United States. Neither finer boundaries nor a more detailed basemap alone resolves such inferential limitations.\n\nMovement displays differ in what they treat as observable. Flow maps connect origins and destinations, often scaling line width to passenger numbers, shipments, or migration counts; a connection does not necessarily identify the route actually traveled. Trajectory displays connect successive recorded positions of an individual object. Their accuracy depends on sampling frequency: a straight segment between two hourly locations may conceal stops or detours. In a space–time cube, two axes represent geographic position and the third represents time, making movement and duration visible as paths but introducing occlusion and perspective effects. Temporal map sequences and small multiples retain a familiar planar geography; fixed extents, classifications, and legends permit comparison across dates. Animation can communicate movement, whereas simultaneous panels allow inspection of earlier states without relying on memory.\n\nInterpolation and spatial density estimation create surfaces from discrete observations, but answer different questions. Inverse-distance weighting estimates a value at an unmeasured location by weighting nearby measurements, commonly using \\(w_i=d_i^{-p}\\); the exponent \\(p\\) controls how strongly nearby observations dominate. Kriging uses a model of spatial covariance to calculate predictions and associated prediction uncertainty. Both depend on assumptions about spatial continuity, sampling, and barriers. Kernel density estimation instead distributes each event’s weight over a neighborhood to estimate spatial concentration. Its bandwidth determines smoothing: small bandwidths reveal local peaks, while large ones merge them. A crime-event density surface represents recorded incidents per area, not automatically individual risk, which also requires an appropriate population or exposure denominator. Measured sites and inferred surfaces therefore represent different kinds of evidence."
    },
    {
      "heading": "Networks, Trees, and Relational Data",
      "text": "Relational displays represent entities and their connections using structures formalized in graph theory. In a node–link diagram, marks represent vertices and connecting lines represent edges; arrows distinguish direction, while line width may encode weight. This format supports following short paths and recognizing branching structures, although crossings can make individual connections difficult to trace. An adjacency matrix instead assigns vertices to rows and columns, with each cell indicating whether, or how strongly, its corresponding pair is connected. An undirected network produces a symmetric matrix when both axes use the same ordering. Matrices avoid edge crossings and can expose dense groups as blocks after suitable reordering, but following a multistep path requires repeated movements between rows and columns. Node–link diagrams often suit sparse networks and path-following tasks; matrices become useful for inspecting dense connectivity, though their effectiveness depends on ordering and the task.\n\nNetwork layouts determine which structures become visually prominent. Force-directed methods model attraction along edges and repulsion between vertices; Thomas Fruchterman and Edward Reingold’s 1991 algorithm is a widely used example. Such methods can produce spatial groupings, but apparent proximity is an outcome of an optimization procedure, not necessarily a measured physical distance or a direct measure of relationship strength. Layered layouts place vertices into successive ranks, making them useful for dependencies and other directed graphs. The framework published by Kozo Sugiyama, Shojiro Tagawa, and Mitsuhiko Toda in 1981 separates the problem into stages including layer assignment, crossing reduction, and coordinate assignment. Circular layouts emphasize ordering or connections around a ring, while geographic layouts anchor vertices to known locations. Each imposes different constraints on the interpretation of position.\n\nHierarchical displays distinguish explicit parent–child connections from spatial containment. Node–link trees show ancestry through connecting edges, with depth commonly encoded by horizontal or vertical position. Dendrograms used in hierarchical clustering show successive merges or splits; when branch height represents a linkage distance, that scale carries information beyond membership. Leaf order is generally not unique, because branches can be rotated without changing the hierarchy. A treemap, introduced by Ben Shneiderman in 1991, recursively partitions a rectangle into nested regions. Leaf area can represent an additive quantity, such as file size, while enclosing regions identify directories. A sunburst diagram uses concentric rings for depth and angular sectors for branches. Both containment approaches use space compactly, but small regions can become difficult to label or compare. Equal-area leaves and quantity-weighted leaves answer different questions and require explicit identification.\n\nA Sankey diagram represents quantities moving between sources, processes, or destinations through bands whose widths are proportional to values. In an energy balance, a transformation stage can divide incoming energy into useful output and losses; losses must remain represented if the diagram is to preserve the accounting balance. Conservation is therefore a property of the underlying quantities and system boundary, not something guaranteed by the visual format. Alluvial plots connect category memberships across successive stages, for example tracking the same students between enrollment statuses in consecutive years. Interpreting these bands as transitions requires linked observations or another justified allocation method: separate category totals alone do not identify individual movements. Category definitions, denominators, and treatment of entrants, exits, and missing records must remain consistent or be explicitly reconciled.\n\nRelational displays face problems beyond the choice of layout. Edge crossings obscure continuity, overlapping vertices conceal entities, and large nodes or thick links can cover weaker connections. Edge bundling reduces clutter by routing connections together, but can make individual endpoints ambiguous. Group enclosures and color can identify communities, although a spatial cluster is not itself evidence of a formally detected community. Layout instability is particularly consequential when comparing networks over time: rerunning an optimization can move unchanged vertices and create an impression of structural change. Fixed coordinates, shared initialization, or constraints on vertex movement can preserve correspondence, at the cost of allowing less freedom to optimize each individual view."
    },
    {
      "heading": "Scientific Fields and Three-Dimensional Data",
      "text": "Scientific datasets often describe quantities distributed through physical space and time. A scalar field assigns one value to each location, such as temperature in a combustion simulation or X-ray attenuation reconstructed by computed tomography. A vector field assigns a magnitude and direction, as in measured fluid velocity or a simulated magnetic field. A tensor field represents a more complex local relationship: a three-dimensional stress tensor describes directional mechanical stresses, while diffusion tensor imaging research models water diffusion using a symmetric \\(3 \\times 3\\) tensor. Tensor glyphs may use ellipsoid axes to show eigenvector directions and axis lengths to encode eigenvalues or their transformations; their appearance therefore depends on the encoding, not simply on a directly observed object. These fields may be sampled on regular voxel grids, irregular meshes, or scattered measurement locations.\n\nSlices, contours, and isosurfaces expose different subsets of a field. A slice samples values on a plane or curved surface; axial, coronal, and sagittal sections are standard orientations for volumetric anatomical images. A contour connects locations with the same scalar value in two dimensions, while an isosurface represents the three-dimensional set \\(f(x,y,z)=c\\). William Lorensen and Harvey Cline’s marching cubes algorithm, published in 1987, constructs triangular approximations to isosurfaces by examining values at the eight corners of grid cells. The selected threshold controls which structures appear, and the resulting surface does not display all values in the volume. Direct volume rendering instead accumulates contributions along viewing rays. A transfer function maps sampled values, sometimes together with gradient magnitude, to color and opacity. Low opacity exposes deeper structures but mixes contributions from multiple depths; high opacity can conceal them. Transfer-function settings are therefore part of the substantive representation, rather than merely cosmetic choices.\n\nVector fields require representations of direction as well as magnitude. Arrow glyphs encode direction through orientation and may encode speed through length or color, but dense placement causes overlap and sparse placement can miss small features. Streamlines follow the instantaneous direction of a field at a fixed time \\(t_0\\), satisfying \\(d\\mathbf{x}/ds=\\mathbf{v}(\\mathbf{x},t_0)\\). Pathlines follow particles through a changing field, satisfying \\(d\\mathbf{x}/dt=\\mathbf{v}(\\mathbf{x},t)\\). Their geometric trajectories coincide in a steady flow but generally differ in an unsteady one. Particle tracing numerically integrates such trajectories from seed locations; seed placement, integration step size, interpolation, and termination criteria all affect the displayed result. Animated tracers can reveal transport, but their apparent density also reflects how particles were introduced, not necessarily fluid density.\n\nSpatial discretization and rendering introduce additional interpretive choices. Mesh cells define where values are stored and how values between samples are reconstructed: trilinear interpolation is common within regular grid cells, while tetrahedral meshes often use piecewise-linear interpolation. Interpolation estimates intermediate values without recovering unresolved physical detail. Coordinate transformations must preserve the meaning of geometry and directional quantities: rotating positions without consistently transforming vectors or tensors produces incompatible representations. Perspective, shading, silhouettes, and depth-dependent attenuation help establish shape, but illumination can introduce brightness differences unrelated to measured values. Opaque surfaces occlude interior structures, and transparent surfaces can make depth ordering ambiguous. Clipping planes and cutaway views expose otherwise hidden regions while also removing context.\n\nA two-dimensional section can outperform an interactive three-dimensional scene when the task is to compare values at known locations, inspect a boundary, or measure a cross-section. It eliminates foreground occlusion and permits stable side-by-side comparisons at matched coordinates. Projections compress an entire volume but require an explicit reduction rule: maximum-intensity projection retains the largest sampled value along each ray, whereas averaging or summation answers different questions. Structures at different depths may overlap in any projection. Three-dimensional views are useful for spatial relationships, yet rotation requires viewers to integrate successive views mentally; linked sections and an overview can support localization without making every comparison depend on manipulating the scene."
    },
    {
      "heading": "Statistical Inference and Uncertainty",
      "text": "Variability among observations is distinct from uncertainty about a quantity inferred from them. A distribution of recorded temperatures describes differences across observations; uncertainty about its population mean depends on sampling design, sample size, and other assumptions. For independent observations with common finite variance, the estimated standard error of the mean is \\(s/\\sqrt{n}\\), whereas the standard deviation \\(s\\) describes observational spread. Increasing sample size can reduce sampling uncertainty without removing calibration bias or measurement uncertainty shared across observations. Model choice, missing-value imputation, classification, and aggregation introduce further uncertainty that a sampling-based interval may not include.\n\nError bars are a graphical form, not a statistical definition: they may indicate standard deviations, standard errors, confidence intervals, or another range, so captions must identify their meaning and coverage. A 95% confidence interval comes from a procedure that, under its assumptions, covers the fixed parameter in 95% of repeated samples; it does not assign a 95% probability to that parameter lying inside the particular observed interval. A 95% credible interval contains 95% of the posterior probability under a specified likelihood and prior. A prediction interval concerns a future observation and usually includes both uncertainty about the fitted mean and observation-level variation, making it wider than a corresponding interval for the mean. These interpretations are not interchangeable.\n\nUncertainty bands extend intervals across a curve, surface, or time sequence. A pointwise 95% band provides coverage at each fixed location separately, not a 95% probability that the entire true curve lies within it. Simultaneous bands address coverage across a specified domain and generally require wider limits. Shaded ribbons can also conceal asymmetry, multiple modes, or extreme tails in a distribution. Ensemble displays instead show multiple plausible curves or model realizations, exposing alternative trajectories, although overlapping lines can obscure density. Ensemble spread need not represent calibrated probability, particularly when models share assumptions. Quantile dot plots allocate equal probability mass to individual dots: with 20 equally weighted dots, each represents 5% of the displayed distribution. Hypothetical outcome plots, introduced under that name by Jessica Hullman, Paul Resnick, and Eytan Adar in 2015, animate successive draws from a distribution. Joint draws can preserve dependence between quantities that separate error bars would conceal.\n\nGraphical diagnostics examine whether fitted models reproduce relevant features of observations. For linear regression, plotting residuals \\(e_i=y_i-\\hat y_i\\) against fitted values can reveal curvature inconsistent with a linear mean function or a widening spread inconsistent with constant error variance. Residuals plotted in observation order can expose temporal dependence. A quantile–quantile plot compares ordered observations or residuals with theoretical quantiles, or with quantiles from another sample; systematic departures from a reference line indicate distributional differences, with tail departures particularly relevant to extreme-value assumptions. Posterior predictive displays compare observed data with replicated data generated from the fitted Bayesian model. Discrepancies in zero counts, maximum values, or group-level variation can reveal aspects the model fails to reproduce, without establishing that a model passing those checks is correct.\n\nVisible patterns can arise from random variation, especially when many panels, variables, or subgroup comparisons are inspected. Under independence, 20 tests conducted at the 5% significance level have probability \\(1-0.95^{20}\\approx0.64\\) of producing at least one false rejection when all null hypotheses hold. Dependence changes this calculation, but the example illustrates why multiple comparisons matter in graphical exploration. Confounding can produce an association without the proposed causal mechanism. In Simpson’s paradox, discussed by Edward H. Simpson in 1951, an association can reverse after groups are combined because group composition and baseline outcomes differ. Showing subgroup estimates, sample sizes, and aggregation weights makes the reversal inspectable; neither the pooled nor the stratified display alone determines the appropriate causal interpretation."
    },
    {
      "heading": "Interaction and Exploratory Interfaces",
      "text": "Interactive visualization connects user actions to changes in visual representations, query results, or both. Selection identifies records or marks for further operations without necessarily removing anything; filtering restricts which records are included. Sorting changes an ordering, potentially revealing ranks or groups while retaining the same observations. Zooming changes the visible scale or extent, whereas panning moves the viewport without changing its scale. Details-on-demand exposes information omitted from the overview, such as a record’s exact values, metadata, or associated document. These operations have different analytical consequences: filtering a sales display to one region changes the population shown, while zooming into a geographic region may merely change the viewport. Ben Shneiderman’s 1996 formulation, “Overview first, zoom and filter, then details-on-demand,” described a useful interaction pattern rather than a mandatory sequence for every interface.\n\nBrushing and linking coordinates selection across multiple views. A user might drag a rectangular brush over points in a scatterplot of income against expenditure; the corresponding records are then highlighted in an age histogram and a geographic display. Linking normally depends on shared record identifiers or explicit relationships between datasets, not on marks occupying similar screen positions. In aggregated views, a selected set may appear as a highlighted portion of each bar rather than as individually identifiable marks. Interfaces must distinguish whether selections replace, extend, or intersect earlier selections, and whether unselected records remain visible. Keeping them visible in muted form can preserve the comparison population; removing them instead converts selection into filtering.\n\nDynamic queries couple controls such as sliders, checkboxes, and range selectors to rapidly updated results. Christopher Ahlberg, Christopher Williamson, and Ben Shneiderman’s 1992 HomeFinder prototype demonstrated this approach for exploring housing through constraints including price and number of bedrooms. Immediate visual updates let users investigate trade-offs by moving controls rather than repeatedly composing and submitting queries. Drill-down, by contrast, moves through defined levels of detail—for example, from annual revenue to quarters, months, and individual transactions. Its interpretation depends on aggregation rules: a revenue total can be decomposed by summation, whereas a percentage generally requires its underlying numerator and denominator to be recomputed.\n\nFocus-plus-context techniques allocate more visual space or detail to an area of interest while retaining surrounding structure. George Furnas’s 1986 generalized fisheye framework formalized a degree-of-interest approach combining an item’s intrinsic importance with its distance from the current focus. Distorted views and overview-plus-detail arrangements address the difficulty of inspecting local information without losing orientation, although distortion can complicate judgments of distance and size. Semantic zooming changes the representation itself as scale changes: a document collection might show category labels at a distant view, document titles closer in, and text excerpts at greater magnification. Unlike simple geometric enlargement, it requires rules governing which information appears at each level.\n\nExploratory work also requires management of analytical state. Interaction histories record actions or visited states; undo reverses operations; bookmarks preserve meaningful configurations without requiring every intermediate action to be replayed. A useful saved state can include filters, selections, sorting, viewport, transformations, and the dataset version. These distinctions matter when collaborators revisit an apparent anomaly: an image preserves appearance, but not necessarily the query or data needed to reconstruct it. Branching histories can retain alternative lines of investigation rather than treating exploration as a single linear sequence.\n\nFeedback and timing determine whether these controls remain intelligible. Selected marks, active-filter summaries, loading indicators, and explicit empty-result messages distinguish successful actions from failures or unfinished computation. Delayed queries can produce stale responses unless the system cancels superseded requests or associates results with the state that generated them. Discoverable controls expose possible actions instead of relying entirely on undocumented gestures. Animated transitions can preserve object identity and show how reordered or regrouped marks move between states, but simultaneous movement, disappearing marks, and long transitions can impede tracking. Stable layouts, consistent encodings, and clear endpoints help preserve the user’s mental model when animation alone cannot explain a change."
    },
    {
      "heading": "Explanatory Graphics, Dashboards, and Data Storytelling",
      "text": "Exploratory interfaces support questions that may change during analysis; explanatory graphics organize evidence around a specified question, audience, and intended interpretation. The distinction concerns purpose rather than whether a display is interactive: a static scatterplot can support exploration, while an interactive article can guide readers toward a carefully delimited finding. Explanatory design selects relevant variables, comparisons, and detail without necessarily showing every analytical step. In research communication, for example, a figure comparing treatment groups may display observations, an estimated difference, and a confidence interval, while its caption identifies the population and explains the interval. The explanatory claim must remain distinguishable from the observations and statistical assumptions supporting it.\n\nAnnotation connects visible patterns to their significance. Direct labels identify series without repeated legend lookup; reference lines mark thresholds or baselines; callouts connect a particular observation to an event or methodological qualification. Contextual comparisons establish whether a value is unusual relative to a historical range, a peer group, or an explicit target. These comparisons are not interchangeable: exceeding last year’s value does not necessarily mean meeting a policy target. Sequencing can introduce the overall distribution before highlighting a subgroup, or establish a baseline before showing change. In journalism, this creates a visual narrative whose sequence determines when readers encounter evidence and qualifications. Edward Segel and Jeffrey Heer’s 2010 study of narrative visualization described structures including the “martini glass”: an author-directed introduction followed by a more open-ended interactive stage.\n\nA dashboard brings several indicators together to support recurrent observation of a system. Coordination requires consistent units, time periods, category definitions, and filter states, not simply placing charts on one screen. Each indicator needs a reporting cadence and a distinction between the period measured and the time of the latest update. A hypothetical weekly service dashboard might report the percentage of requests completed within 48 hours against a 90% target; its definition must specify eligible requests, whether weekends count, and how unresolved cases enter the denominator. Otherwise, identical labels can conceal different measurements. Monitoring identifies departures from expected conditions; diagnosis investigates their causes. A falling completion rate can trigger investigation, but identifying whether staffing, demand, or recording changes produced it requires additional evidence. Linked breakdowns can support that investigation without themselves establishing causation.\n\nInteractive articles combine authored explanation with reader-controlled inspection. Scrollytelling uses scrolling to advance explanatory stages, often retaining a chart while changing its annotations, highlighted subset, or scale. A step-by-step sequence can introduce unfamiliar encodings before presenting a complex comparison, whereas filters and details-on-demand allow readers to examine cases beyond the main account. These controls distribute authority differently: the author determines the initial framing, while the reader chooses some subsequent questions. Animated explanations add temporal ordering to this arrangement. Hans Rosling’s 2006 TED presentation used animated country-level bubbles to relate income, life expectancy, population, and historical change. Such animation can make trajectories salient, but exact comparisons between distant moments may still require static frames or persistent reference marks.\n\nCaptions and publication context make an explanation interpretable outside its original layout. A self-contained figure identifies what is measured, the population and geographic coverage, units, relevant dates, and the meaning of symbols or intervals. Source notes distinguish original data providers from subsequent calculations and specify consequential transformations, such as inflation adjustment or seasonal adjustment. Definitions clarify whether “growth” means an absolute increase, percentage change, or change in percentage points. Publication and revision dates matter particularly for frequently updated indicators: a screenshot may otherwise appear current after its underlying dataset has changed. Headlines, nearby prose, and figure captions also need consistent scopes, so that a finding about one subgroup or reporting period is not presented as a claim about an entire population."
    },
    {
      "heading": "Accessibility and Visualization Literacy",
      "text": "Accessibility concerns whether people can perceive, operate, and interpret a visualization through the interfaces available to them. Color-vision differences can make categories or diverging scales difficult to distinguish, particularly when hue is the only distinguishing feature. Low vision can make small labels, thin lines, and faint gridlines illegible; blindness can prevent access to relationships encoded only spatially. Motor impairments can make precise pointing, dragging, or closely spaced controls difficult to operate. Cognitive demands arise from unfamiliar symbols, dense legends, divided attention, and the need to remember values across views. These barriers interact: magnifying a chart may improve label visibility while hiding the legend or comparison baseline outside the viewport.\n\nAccessible visual encoding uses redundant distinctions, such as directly labeled lines with different dash patterns or symbols as well as colors. The Web Content Accessibility Guidelines (WCAG) 2.2, published as a W3C Recommendation in 2023, specify a minimum text contrast ratio of 4.5:1 for ordinary text and 3:1 for large text at Level AA, with defined exceptions. Required graphical objects and visual information identifying interface components or states generally require 3:1 contrast against adjacent colors. These thresholds do not establish that every mark in a dense chart is readily distinguishable. Scalable text, sufficient spacing, and layouts that remain usable under magnification address additional barriers. Keyboard users need a logical navigation order, visible focus, and ways to perform selection, filtering, and inspection without a mouse. Information exposed by hovering over a point also needs an alternative, such as keyboard focus or a persistent details panel; a chart containing thousands of points need not place every point in the tab sequence.\n\nNonvisual access requires more than naming the chart type. A short text alternative can identify its subject and principal finding, while a longer description explains axes, units, coverage, major patterns, and relevant exceptions. For example, a description of a monthly unemployment chart can state the period shown, the location and magnitude of its peak, and whether the series subsequently declined. A structured data table with properly associated row and column headers supports exact-value lookup through screen-reader navigation, but does not automatically communicate a trend that is readily visible in a line chart. Sonification maps data to sound—for example, encoding increasing values as rising pitch while time progresses through playback. Tactile graphics use raised lines, textures, symbols, and Braille labels to communicate spatial relationships. Both approaches require explicit mapping conventions; converting every visual detail directly into sound or touch can create clutter rather than equivalent access.\n\nVisualization literacy includes the ability to read, interpret, question, and construct graphical representations. It involves recognizing encodings, comparing quantities, understanding transformations and uncertainty, and questioning denominators or omitted observations. Unfamiliarity with a format is distinct from lack of domain knowledge: an experienced clinician may understand the measurements in a parallel-coordinates plot without knowing how its crossing lines encode records. Conversely, recognizing a familiar bar chart does not establish an understanding of the population or sampling procedure behind it. Assessment instruments such as the Visualization Literacy Assessment Test, introduced in 2017, use chart-based questions to examine interpretation skills. Performance on such tasks measures particular reading abilities, not every aspect of critical judgment or graphic construction.\n\nLocalization affects interpretation as well as interface text. The date “03/04/2025” can denote 3 April or March 4, and decimal and thousands separators differ across locales. Translated labels may require more space; units, currencies, calendar conventions, and time zones need explicit treatment. Right-to-left interfaces require attention to navigation and label placement without mechanically reversing axes whose ordering has substantive meaning. Accessibility testing therefore combines standards checks with tasks performed by intended users, including people using screen readers, magnification, alternative input devices, and different language settings. Such testing can reveal failures that contrast measurements or automated checks miss, including inaccessible filter states and descriptions that omit the comparison users need."
    },
    {
      "heading": "Software, Specifications, and Rendering",
      "text": "Spreadsheets combine cell-based calculation, tabular inspection, and chart creation, making graphics closely dependent on worksheet ranges, formulas, and refresh behavior. Statistical environments such as R and Python integrate visualization with data transformation, model fitting, and scripted analysis. Business-intelligence systems, including Tableau and Microsoft Power BI, connect reports to data sources and shared semantic models that define measures, relationships, and aggregation rules. A geographic information system, such as QGIS or ArcGIS, additionally manages coordinate reference systems, spatial joins, and geographic layers. Specialized scientific tools address meshes, fields, and volumes: ParaView provides an application interface and processing pipelines built on VTK, the Visualization Toolkit. These categories overlap; their differences concern supported data models, workflows, and deployment rather than mutually exclusive chart types.\n\nA grammar of graphics describes plots through composable elements rather than a separate procedure for every named chart. Hadley Wickham’s ggplot2, first released in 2007, implements a layered grammar in R: data variables map to aesthetics, geometric objects draw marks, statistical transformations derive quantities, and scales translate values into visual properties. For example, the same category and measurement mappings can support raw points and an additional layer of group summaries. Vega expresses visualizations as declarative JSON specifications containing data transformations, scales, marks, and interactive signals. Vega-Lite uses a higher-level specification of fields, data types, marks, and encodings, which compiles into Vega; its 2017 research paper described a grammar supporting interactive graphics. Declarative specifications state relationships and intended behavior while a runtime handles many execution details.\n\nMore explicitly programmed approaches expose different implementation choices. Matplotlib, begun by John Hunter in 2003, supplies Python plotting interfaces organized around objects such as figures, axes, and artists; it also offers higher-level convenience functions. D3, introduced by Michael Bostock, Vadim Ogievetsky, and Jeffrey Heer in 2011, uses JavaScript to bind data to document elements and compute attributes, styles, and transitions. A D3 data join can associate records with graphical elements through stable keys, distinguishing entering, updating, and exiting records. Neither library implies unrestricted low-level programming: both support reusable abstractions. Conversely, declarative systems commonly expose application programming interfaces for embedding graphics, supplying data, and responding to selections.\n\nOutput formats determine how graphical information is stored. Raster images represent a fixed grid of pixels; enlarging a PNG beyond its native dimensions exposes that finite resolution. Vector formats represent shapes, paths, and text, allowing many chart elements to remain sharp when resized. PDF and Scalable Vector Graphics (SVG) can preserve such geometry, although files may also contain embedded raster images. Vector output is not necessarily smaller: a scatterplot containing hundreds of thousands of separate marks can produce a large document. Export also depends on fonts, clipping, transparency, and color handling; a visually correct screen display does not guarantee identical appearance in print.\n\nWithin a web browser, SVG creates document elements that can receive styles and events individually. Canvas instead exposes a drawing surface: applications issue drawing commands and generally maintain their own representation of objects for redrawing and hit testing. WebGL provides a graphics interface using the graphics processing unit, enabling shader-based rendering of two- and three-dimensional scenes. These are rendering mechanisms, not chart-design methods. SVG often supports direct inspection of individual marks, while Canvas and WebGL can avoid maintaining a document element for every mark. Their practical performance depends on geometry, updates, interaction logic, and implementation. Exporting a Canvas or WebGL image normally captures pixels unless the application separately reconstructs vector geometry.\n\nReliable production also requires reusable components and explicit dependencies. Shared axis, legend, and formatting components can enforce consistent behavior across reports, while versioned schemas and interfaces help prevent silent changes in field interpretation. Tests can check transformed values, scale domains, category ordering, interaction states, and rendered images; image comparisons require tolerances for platform-dependent font rendering. Reproducible generation records input versions, transformation code, software dependencies, and configuration. Randomized layouts require controlled seeds, and interactive publications may additionally require saved filters and selection states to reproduce the displayed result."
    },
    {
      "heading": "Large-Scale, Streaming, and Distributed Data",
      "text": "Large datasets encounter several independent limits before their records become visible. A 1920 × 1080 display contains 2,073,600 pixels, so plotting ten million points necessarily creates substantial overlap even before accounting for axes and labels. Memory requirements include not only source values but also indexes, intermediate results, and rendering buffers: ten million pairs of 64-bit coordinates alone occupy 160 million bytes, approximately 153 MiB. Query execution, network transfer, layout computation, and drawing impose separate costs. A fast renderer cannot compensate for a query that scans billions of records after every interaction; conversely, a fast query does not make millions of individually interactive graphical objects inexpensive to maintain.\n\nReduction methods address different parts of this problem. Random sampling reduces record count but can miss rare events or small subpopulations; stratified sampling preserves specified groups, although population-level summaries then require appropriate weighting. Binning replaces individual observations with counts or statistics over intervals or spatial cells, hiding within-bin variation and making boundaries consequential. Aggregation can retain counts, sums, and extrema, but averages alone conceal distributions. Level-of-detail methods change representation with scale—for example, showing regional totals before revealing individual locations when zooming. Multiresolution pyramids store several such scales in advance. The 2013 systems imMens and Nanocubes demonstrated complementary approaches: imMens used precomputed, binned data tiles and GPU processing for interactive views, while Nanocubes used a compact hierarchical index for querying large spatiotemporal datasets. Neither approach makes every original observation individually visible at every scale.\n\nDatabase-backed visualization moves filtering and aggregation toward the stored data rather than transferring complete tables to a client. Spatial indexes prune irrelevant regions, column-oriented storage reduces access to unused attributes, and cached query results avoid repeated work. Prefetching can prepare likely next views, although predictions consume resources and cached results need invalidation when data change. Parallel processing divides scans or aggregations across processors; distributed execution extends this across machines but introduces communication, coordination, and data-partitioning costs. A GPU can accelerate both numerical operations and rendering, particularly when many records undergo the same operation. Transfers between main memory and GPU memory, however, can become a bottleneck. Performance therefore depends on the whole interaction pipeline, not simply the number of marks drawn per second.\n\nContinuously updated displays add time and ordering constraints. Stream processing commonly uses tumbling windows, which do not overlap, or sliding windows, which do: a ten-minute window refreshed every minute reuses nine minutes of observations. Incremental computation updates stored summaries instead of recomputing them from scratch; a running mean, for example, can be maintained using a sum and count. Event time records when an observation occurred, whereas processing time records when the system handles it. Late arrivals can consequently revise already displayed totals. Watermarks track estimated event-time progress under a system’s lateness policy; they do not establish that every earlier event has arrived. Backpressure slows upstream production when downstream stages cannot keep up. Buffering, coalescing updates, or dropping records are alternative responses, each with different consequences for delay and completeness.\n\nProgressive visualization presents useful intermediate results while computation continues, allowing exploration before a full query finishes. Partial scans, successive samples, or increasingly fine spatial resolutions can provide early feedback, but their meanings differ: the first partitions scanned may be geographically or temporally biased, whereas a probability sample supports particular statistical estimates. Interfaces can distinguish these states through processed-record counts, data timestamps, refinement indicators, and clearly labeled approximate values. Where justified by the estimator, confidence intervals can communicate sampling uncertainty, but they do not account automatically for missing sources or systematic bias. Rankings, clusters, and apparent outliers may change during refinement; preserving the user’s selection and identifying revised values helps prevent an intermediate result from being mistaken for a stable finding."
    },
    {
      "heading": "Evaluation and Research Methods",
      "text": "Evaluation addresses several distinct outcomes rather than a single measure of visualization quality. Encoding accuracy concerns how faithfully people estimate represented quantities; task performance concerns success in activities such as finding anomalies or comparing groups. Usability includes learnability, interaction difficulties, and recovery from mistakes. Evaluating analytical reasoning requires examining how people develop, test, and revise explanations, not merely counting correct answers. Adoption measures continued use, whereas organizational usefulness concerns consequences such as reduced reporting effort or better coordination. These outcomes can diverge: a frequently used dashboard may be required by management, and a perceptually accurate chart may not answer the questions that matter in a workflow. Heidi Lam and colleagues’ 2012 paper, Empirical Studies in Information Visualization: Seven Scenarios, organized evaluation around differing purposes, including user performance, user experience, and visual data analysis and reasoning.\n\nControlled experiments isolate particular design choices by comparing conditions under specified tasks. A study might compare two uncertainty displays while holding the underlying distributions, questions, and screen dimensions constant. Measures include response accuracy, completion time, numerical estimation error, and error patterns, such as consistently confusing a distribution’s spread with uncertainty about its mean. Within-participant designs expose each person to multiple conditions but require counterbalancing to limit order and learning effects; between-participant designs reduce carryover while introducing differences between participant groups. Experimental design must also distinguish participants from their repeated trials: hundreds of responses from a few people do not constitute hundreds of independent participants. Effect sizes and confidence intervals help characterize differences beyond whether a statistical test crosses a significance threshold.\n\nParticipant expertise and task selection determine what such results support. Identifying the largest bar tests a different capability from investigating an unexpected change in an operational dataset. Domain specialists may exploit terminology and contextual knowledge unavailable to novices, while experienced visualization users may recognize unfamiliar encodings more readily. Jeffrey Heer and Michael Bostock’s 2010 study, Crowdsourcing Graphical Perception, demonstrated the use of Amazon Mechanical Turk for graphical-perception experiments, including replication of earlier laboratory findings. This established a practical recruitment approach, not a guarantee that online samples represent every intended audience. Device characteristics, viewing conditions, participant screening, and data-quality checks remain consequential.\n\nQualitative research examines activities that fixed-answer tasks can miss. Observation records how visualizations fit into meetings, handoffs, and existing software practices; interviews elicit goals, interpretations, and perceived limitations. Think-aloud studies ask participants to verbalize their reasoning during use, revealing misunderstood labels or unexpected interaction strategies, although verbalization can alter pace and behavior. Longitudinal field studies can show how learning, trust, and use change over weeks or months. Interaction logs document actions but do not, by themselves, establish why an action occurred or whether it produced understanding. Combining logs, observations, interviews, and analysis artifacts provides complementary evidence.\n\nDesign studies investigate visualization through sustained collaboration on a concrete domain problem. Michael Sedlmair, Miriah Meyer, and Tamara Munzner’s 2012 methodological paper described a nine-stage framework extending from preparation and problem characterization through deployment and reflection. Such studies can contribute transferable problem abstractions and design lessons alongside a working system. Benchmark datasets serve a different purpose: shared inputs and tasks enable comparisons under common conditions, although a benchmark’s coverage limits the conclusions. Reproducibility materials can include data provenance, preprocessing scripts, software versions, stimuli, task instructions, anonymized responses, and analysis code.\n\nThreats to validity arise throughout this process. Practice, fatigue, and condition order can distort comparisons; small or unrepresentative samples constrain precision and generalization. Artificial tasks may reward rapid lookup while excluding the ambiguous questioning central to actual analysis. Novelty can temporarily increase engagement, and deployment outcomes may reflect training, staffing, or policy changes rather than the visualization alone. Preregistered hypotheses, transparent exclusions, counterbalanced conditions, and evaluation across laboratory and field settings address different weaknesses; none independently establishes that a finding transfers to other users, datasets, or organizations."
    },
    {
      "heading": "Integrity, Ethics, and Data Governance",
      "text": "A visualization can mislead through technically correct numbers combined with inappropriate graphical choices. Truncating a bar-chart axis at 95 makes values of 98 and 100 appear in a height ratio of 3:5, although their numerical ratio is 0.98; because bar length encodes magnitude, the displaced baseline changes the comparison. A nonzero baseline in a line chart is not inherently misleading when the purpose is to examine small fluctuations and the scale is explicit. Three-dimensional extrusion can introduce perspective distortion and occlusion without adding data. Selective time windows can make a temporary rise appear to be a sustained trend, while omitted denominators can turn differences in population size into apparent differences in risk. Counts, rates, and absolute versus relative changes answer different questions. These features establish grounds for criticizing a representation, but do not by themselves establish deliberate deception.\n\nCategories and exclusions determine whose experiences become visible. Administrative records describe people encountered by an institution, not necessarily everyone in the population of interest: service-use records, for example, omit people who need a service but cannot access it. A category labelled “other” can combine groups with substantially different outcomes, while complete-case analysis can exclude populations with higher levels of missing information. Institutional definitions also change over time. The United States’ 2000 population census first allowed respondents to select more than one race, complicating direct comparisons with earlier single-race tabulations. Graphical continuity across such changes can conceal discontinuity in measurement. Definitions, coverage, exclusions, and category revisions therefore form part of a graphic’s substantive meaning, rather than merely its documentation.\n\nPrivacy risks arise when displays expose precise locations, rare combinations of attributes, or small groups. Removing names does not necessarily prevent identification: in 2008, Arvind Narayanan and Vitaly Shmatikov demonstrated how background knowledge, including publicly available IMDb ratings, could help identify users in the Netflix Prize’s anonymized movie-rating dataset. Interactive filters can create similar vulnerabilities by allowing successive queries to isolate individuals or infer suppressed values through subtraction. Geographic aggregation, suppression, rounding, and restricted query access reduce particular disclosure risks, but can also hide local variation or reduce the visibility of small populations. Suppressing one table cell may be ineffective if its value can be reconstructed from published totals.\n\nDifferential privacy formalizes a different approach: a randomized release mechanism limits how much its output distribution can change when one person’s data are added or removed. Its privacy parameter governs a mathematical bound, not a guarantee that every displayed statistic is accurate. The United States Census Bureau used differential privacy in producing 2020 Census redistricting data. Such disclosure controls create trade-offs among privacy, accuracy, geographic detail, and consistency across tables. Small-area graphics can be especially sensitive to the introduced error. Descriptions of protected data need to distinguish published estimates from unmodified observations and identify relevant limitations without exposing protected records.\n\nData governance establishes responsibility for acquisition, transformation, authorization, publication, and correction. Source attribution identifies an origin; copyright and licensing determine whether particular reuse is permitted. Neither attribution nor a license alone establishes data quality. Data provenance records how an output derives from inputs, including dataset versions, joins, exclusions, calculations, and responsible agents. The W3C’s 2013 PROV recommendations provide a model organized around entities, activities, and agents. For a published graphic, inspectable provenance can include the retrieval date, transformation code, units, denominator definitions, and revision history. Corrections are more traceable when notices identify what changed, when it changed, and whether the interpretation changed, rather than silently replacing the image.\n\nEvidence-based interpretation and persuasive framing are not opposites: choosing a comparison, headline, or annotation inevitably directs attention. The ethical distinction concerns whether the resulting claim is supported and whether consequential choices remain inspectable. A graphic showing improvement against a selected baseline may support that comparison without supporting a claim that a policy caused the improvement; confounding and alternative explanations remain relevant. Publishing underlying values, explaining benchmark selection, exposing active filters, and distinguishing observations from projections allow readers to examine the relationship between evidence and assertion."
    },
    {
      "heading": "Applications and Emerging Research",
      "text": "In scientific research, visualization connects measurements and simulations to questions about structure, variation, and model fit: astronomers inspect sky surveys, molecular researchers examine spatial configurations, and climate scientists compare simulated fields with observations. Environmental analysis combines remote sensing, monitoring stations, and geographic information systems to examine vegetation change, atmospheric pollution, and flood exposure. Business applications include inventory monitoring, process control, sales forecasting, and investigation of transaction anomalies; a dashboard may reveal a deviation without identifying its cause. Official statistics use population pyramids, regional maps, and interactive tables to communicate census and economic data. In education, linked graphs and simulations let learners manipulate parameters and observe consequences. Journalism combines graphics with reporting and annotation, while public-health surveillance uses epidemic curves and geographic displays to track reported cases. Johns Hopkins University’s COVID-19 dashboard, launched in January 2020, provided a prominent example of continuously updated geographic reporting; its counts represented reported observations rather than a complete enumeration of infections.\n\nWithin machine-learning workflows, visualization supports inspection of training data, class imbalance, missingness, mislabeled examples, and differences between development and deployment populations. Confusion matrices separate types of classification error, precision–recall curves expose threshold-dependent trade-offs, and calibration plots compare predicted probabilities with observed frequencies. Interactive systems can connect an aggregate error pattern to individual records, allowing analysts to investigate whether failures concentrate in particular conditions or subgroups. Techniques associated with explainable artificial intelligence, including feature-attribution displays and partial-dependence plots, represent selected aspects of model behavior. They do not establish reliability: correlated predictors can complicate attribution, partial dependence can evaluate unrealistic feature combinations, and visually coherent clusters in a projection need not correspond to stable classes. Reliability assessment therefore remains dependent on validation data, appropriate performance measures, and examination of deployment conditions.\n\nAutomated visualization research addresses both chart selection and the translation of analytical intentions into executable specifications. Voyager, introduced in 2015 by Kanit Wongsuphasawat and colleagues, combined manual specification with recommendations for exploring alternative views. Draco, presented in 2018 by Dominik Moritz and colleagues, expressed visualization-design knowledge through constraints, separating prohibited combinations from preferences among acceptable designs. Natural-language systems add the problem of interpreting phrases such as “largest increase,” which may denote absolute change, percentage change, or a comparison over an unspecified interval. Generative AI systems can produce database queries, analysis code, and chart specifications; LIDA, presented by Victor Dibia in 2023, exemplified a language-model-based pipeline for generating visualizations. Failure modes include nonexistent field names, incorrect joins or aggregations, executable queries that answer the wrong question, and graphics whose encodings do not support the requested comparison. Syntactic validation detects only some of these errors; checking meaning requires comparison with the data schema, analytical intent, and computed results.\n\nImmersive analytics investigates how virtual reality, augmented reality, spatial interaction, and shared environments affect analytical work. A three-dimensional scatterplot can place observations around a viewer, while a tracked controller permits selection through pointing or spatial enclosure. Augmented displays can register measurements against physical equipment or locations. Collaborative research examines how groups divide tasks, maintain awareness of others’ selections, and coordinate public displays with private workspaces. Embodied approaches also include data physicalization: physical objects whose geometry or material properties encode data. Yvonne Jansen and colleagues’ 2015 research agenda helped establish this term. Tangible arrangements permit manipulation and tactile inspection, but updating them, representing precise values, and providing equivalent access remain distinct design problems.\n\nOpen questions concern how analytical responsibility is distributed between people and automated systems. A recommendation may accelerate exploration while also narrowing attention toward readily generated patterns. Evaluating this trade-off requires more than counting charts or measuring completion time: studies must distinguish observations from justified findings and assess whether findings support subsequent decisions. Trustworthy automation also requires inspectable transformations, recoverable interaction histories, and explicit treatment of uncertain interpretations. Transfer across domains remains difficult because identical graphical operations can depend on different observational units, sampling processes, and professional conventions; success with a benchmark dataset does not establish usefulness in a new institutional workflow."
    }
  ],
  "links": [
    "statistics",
    "computer-science",
    "causation",
    "accessibility",
    "art",
    "algorithm",
    "cluster-analysis",
    "standard-deviation",
    "correlation",
    "linear-regression",
    "map",
    "world-wide-web",
    "javascript",
    "latitude-and-longitude",
    "database",
    "graph-theory",
    "reproducibility",
    "time-series",
    "cognitive-bias",
    "density-estimation",
    "cumulative-distribution-function",
    "dimensionality-reduction",
    "principal-component-analysis",
    "directed-graph",
    "vector-field",
    "standard-error",
    "measurement-uncertainty",
    "confidence-interval",
    "credible-interval",
    "probability-distribution",
    "posterior-predictive-checking",
    "multiple-testing",
    "confounding",
    "journalism",
    "narrative",
    "geographic-information-system",
    "application-programming-interface",
    "web-browser",
    "graphics-processing-unit",
    "parallel-computing",
    "distributed-computing",
    "experiment",
    "experimental-design",
    "qualitative-research",
    "population-census",
    "data-privacy",
    "data-governance",
    "copyright",
    "remote-sensing",
    "education",
    "public-health-surveillance",
    "machine-learning",
    "training-data",
    "explainable-artificial-intelligence",
    "generative-artificial-intelligence",
    "virtual-reality"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}