{
  "id": "association-rule-learning",
  "title": "Association Rule Learning",
  "lang": "en",
  "source_lang": "en",
  "summary": "Association rule learning discovers interpretable co-occurrence relationships among items or attributes in large datasets.",
  "category": "computer-science",
  "subcategory": "artificial-intelligence",
  "type": "concept",
  "aliases": [
    "Association rule mining",
    "Association analysis"
  ],
  "generated_at": "2026-11-04",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/association-rule-learning/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/association-rule-learning/",
      "json": "https://aiwiki.page/zh/association-rule-learning.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Association rule learning is a family of methods in data mining and machine learning for discovering relationships of the form “when these items occur, those items tend to occur as well.” Its classical setting is a collection of transactions, such as shopping baskets, in which each transaction contains a subset of available items. A rule might state that baskets containing pasta and tomato sauce frequently also contain cheese. The rule describes an observed association, not a logical necessity or a causal relationship.\n\nThe field combines a precisely defined search problem with practical questions about usefulness and interpretation. Algorithms identify combinations that occur sufficiently often, generate directional rules from them, and evaluate those rules using measures such as support, confidence, and lift. The principal computational difficulty is that the number of possible combinations grows exponentially with the number of items. The principal interpretive difficulty is distinguishing informative relationships from coincidences, obvious consequences of common items, and artifacts of data collection.\n\nAssociation rule learning became a major research area in the 1990s through work on retail transaction databases. Its methods now appear in recommendation, industrial monitoring, web usage analysis, and exploratory scientific analysis. Although commonly classified as unsupervised learning, it also supports supervised classification when rules are restricted to predicting a designated class."
    },
    {
      "heading": "Mathematical formulation and transaction representation",
      "text": "Let \\(I=\\{i_1,\\ldots,i_m\\}\\) be a finite set of items, and let \\(D=(T_1,\\ldots,T_N)\\) be a collection of transactions, where each \\(T_j\\subseteq I\\). Transactions are indexed because two distinct transactions may contain exactly the same items. An itemset is any subset of \\(I\\); a \\(k\\)-itemset contains \\(k\\) items. In the classical binary formulation, an item either occurs in a transaction or does not. Its quantity, purchase price, and position within the transaction are ignored unless explicitly encoded.\n\nAn association rule has the form \\(X\\Rightarrow Y\\), where \\(X\\) and \\(Y\\) are nonempty, disjoint itemsets. The left-hand side is the antecedent and the right-hand side is the consequent. A rule with several items in its consequent means that those items occur together, not that any one of them suffices. For example, \\(\\{\\text{bread}\\}\\Rightarrow\\{\\text{milk},\\text{eggs}\\}\\) concerns baskets containing both milk and eggs. It is therefore different from either of the two single-consequent rules.\n\nThe support count of an itemset \\(A\\) is\n\n\\[n(A)=\\left|\\{j:A\\subseteq T_j\\}\\right|,\\]\n\nand its relative support is\n\n\\[\\operatorname{supp}(A)=\\frac{n(A)}{N}.\\]\n\nThe support of \\(X\\Rightarrow Y\\) conventionally means \\(\\operatorname{supp}(X\\cup Y)\\). An itemset is frequent if its support reaches a specified minimum, often written \\(\\mathrm{minsup}\\). Classical rule mining seeks all rules satisfying both a minimum support and a minimum confidence threshold.\n\nThis representation is not restricted to purchases. A row in a relational database can become a transaction whose items are attribute–value pairs, such as “region=north” or “device=mobile.” Such encoding makes arbitrary categorical records accessible to transaction-mining algorithms. It also introduces logical constraints: two mutually exclusive values of the same single-valued attribute cannot coexist in a valid record.\n\nDefining the transaction boundary is part of defining the problem. A purchase basket, a customer’s monthly purchases, and a browsing session produce different associations even when derived from the same underlying activity. Combining records across long intervals increases opportunities for unrelated items to co-occur. Splitting records too narrowly may remove meaningful relationships. Similarly, the denominator \\(N\\) determines the population to which support refers. A rule computed only among customers who made purchases cannot automatically be interpreted as a statement about all visitors."
    },
    {
      "heading": "Support, confidence, and measures of association",
      "text": "The confidence of a rule estimates the conditional probability of its consequent given its antecedent:\n\n\\[\\operatorname{conf}(X\\Rightarrow Y) =\\frac{\\operatorname{supp}(X\\cup Y)} {\\operatorname{supp}(X)}.\\]\n\nIt is defined when the antecedent has positive support. Confidence is directional: \\(X\\Rightarrow Y\\) and \\(Y\\Rightarrow X\\) generally have different values, although their supports are identical. This asymmetry reflects different conditioning populations, not necessarily a direction of influence.\n\nConsider 1,000 baskets, of which 200 contain coffee, 300 contain biscuits, and 120 contain both. The rule coffee \\(\\Rightarrow\\) biscuits has support \\(0.12\\) and confidence \\(0.60\\). Its reverse has confidence \\(0.40\\). Both exceed the unconditional frequency of their respective consequents, but confidence alone does not express that comparison. If biscuits appeared in 90% of all baskets, a rule predicting biscuits with 60% confidence would instead describe a negative association.\n\nInterestingness measures address different aspects of this problem. Lift compares observed joint occurrence with the value expected under independence:\n\n\\[\\operatorname{lift}(X\\Rightarrow Y) =\\frac{\\operatorname{supp}(X\\cup Y)} {\\operatorname{supp}(X)\\operatorname{supp}(Y)} =\\frac{\\operatorname{conf}(X\\Rightarrow Y)} {\\operatorname{supp}(Y)}.\\]\n\nThe coffee–biscuits example has lift \\(2\\). Lift is symmetric despite its rule notation. A value of \\(1\\) indicates empirical independence for the two occurrence events; values above or below \\(1\\) indicate positive or negative association, respectively. High lift does not by itself establish reliability: a very rare joint occurrence can produce a large ratio.\n\nLeverage measures an absolute rather than relative departure from independence:\n\n\\[\\operatorname{leverage}(X\\Rightarrow Y) =\\operatorname{supp}(X\\cup Y) -\\operatorname{supp}(X)\\operatorname{supp}(Y).\\]\n\nIts value in the example is \\(0.06\\), equivalent to 60 more joint occurrences than expected under an independence model. Conviction instead compares expected and observed violations of a directional rule:\n\n\\[\\operatorname{conviction}(X\\Rightarrow Y) =\\frac{1-\\operatorname{supp}(Y)} {1-\\operatorname{conf}(X\\Rightarrow Y)}.\\]\n\nWhen confidence is \\(1\\) and the consequent is not universal, conviction is commonly treated as infinite. Degenerate cases require separate handling.\n\nNo single measure captures frequency, predictive precision, departure from independence, uncertainty, and practical value simultaneously. A common consequent may yield high confidence but little new information; a rare consequent may yield impressive lift but few supporting observations. Rule evaluation therefore often reports several measures alongside the underlying counts. Choosing a ranking measure changes the question being answered, rather than merely changing the presentation of an otherwise fixed notion of quality."
    },
    {
      "heading": "Historical development and the frequent-itemset framework",
      "text": "Modern association rule mining was established by Rakesh Agrawal, Tomasz Imieliński, and Arun Swami in their 1993 paper “Mining Association Rules between Sets of Items in Large Databases.” Their formulation addressed transaction data generated by retail operations and expressed the discovery task through user-specified support and confidence thresholds. The work was associated with IBM research and helped establish pattern discovery as a central problem in database-oriented data mining.\n\nThe original AIS algorithm generated candidate itemsets while scanning transactions. It demonstrated the feasibility of extracting associations from large databases, but its candidate-generation strategy could produce substantial unnecessary work. Agrawal and Ramakrishnan Srikant introduced Apriori in 1994, giving a systematic level-wise method that exploited the mathematical structure of support. Apriori became an influential baseline because its pruning rule was simple, general, and independent of a particular retail application.\n\nA decisive conceptual step was separating rule discovery into two tasks. The first is to enumerate frequent itemsets. The second is to generate rules from each frequent itemset and retain those meeting the confidence threshold. If a rule meets minimum support, then its union \\(X\\cup Y\\) must already be a frequent itemset. Consequently, the first task provides a complete search space for the second. This decomposition also makes frequent-itemset mining useful independently of rules, for example in finding recurring configurations or compact descriptions of transaction collections.\n\nResearch subsequently attacked the cost of frequent-itemset enumeration from several directions. Vertical methods represented each item by the identifiers of transactions containing it. Pattern-growth methods compressed shared transaction prefixes and avoided explicit generation of large candidate collections. FP-growth, presented by Jiawei Han, Jian Pei, and Yiwen Yin in 2000, became a prominent example of the latter approach. Work on closed itemsets and compact rule bases addressed the different problem of representing a large result without listing every frequent combination.\n\nThe field also expanded beyond binary, unweighted transactions. Quantitative rules introduced numerical intervals; multilevel rules incorporated taxonomies; sequential pattern mining incorporated order; and utility-oriented methods incorporated quantities and economic value. These extensions retain aspects of the original search problem while changing which patterns count as meaningful.\n\nAssociation rules are distinct from manually authored expert-system rules. Their empirical strength is estimated from observations, and exceptions are expected unless confidence is exactly one in the observed data. They are also distinct from causal rules: the historical objective was to discover co-occurrence structure efficiently, not to identify the effects of interventions."
    },
    {
      "heading": "Apriori and level-wise candidate pruning",
      "text": "Apriori relies on the downward-closure property of support. If \\(A\\subseteq B\\), every transaction containing \\(B\\) also contains \\(A\\), so\n\n\\[\\operatorname{supp}(B)\\leq \\operatorname{supp}(A).\\]\n\nThus every subset of a frequent itemset is frequent, and every superset of an infrequent itemset is infrequent. This anti-monotonicity permits whole branches of the search space to be discarded without counting their transactions individually. It is a property of support, not a general property of every rule-quality measure.\n\nThe algorithm begins by counting single items and retaining the frequent ones, usually denoted \\(L_1\\). It then constructs candidate pairs from them, counts those pairs, and retains \\(L_2\\). At a general stage, frequent \\(k\\)-itemsets are joined to propose \\((k+1)\\)-itemsets. A proposed candidate is discarded if any of its \\(k\\)-item subsets is absent from \\(L_k\\). The database is scanned to count the remaining candidates, and the procedure ends when no further frequent itemsets exist.\n\nSuppose \\(\\{a,b\\}\\) and \\(\\{a,c\\}\\) are frequent. Their union suggests the candidate \\(\\{a,b,c\\}\\). If \\(\\{b,c\\}\\) is not frequent, however, the candidate is rejected before support counting. It cannot possibly meet the threshold. Efficient implementations arrange items in a fixed order so that joins and duplicate avoidance are systematic. A hash table, a prefix tree, or another indexing structure can organize candidates and accelerate the task of finding which candidates are contained in each transaction.\n\nRule generation follows a related but distinct pruning principle. For a fixed frequent itemset \\(L\\), each nonempty proper subset \\(H\\subset L\\) can serve as a consequent, producing \\(L\\setminus H\\Rightarrow H\\). Its confidence is \\(\\operatorname{supp}(L)/\\operatorname{supp}(L\\setminus H)\\). As \\(H\\) grows, the antecedent shrinks and its support cannot decrease, so confidence cannot increase. Therefore, if a consequent fails the confidence threshold, its supersets can be pruned during rule generation for that same union \\(L\\).\n\nApriori is especially effective when transactions are relatively short and minimum support is high enough to eliminate most combinations early. Its difficulties arise when many items are frequent or long frequent patterns exist. Candidate sets can become enormous, and each level may require another scan. For a frequent itemset containing \\(k\\) items, there are \\(2^k-2\\) possible nonempty directional splits. Efficient itemset enumeration therefore does not remove the separate risk of an enormous rule output."
    },
    {
      "heading": "Pattern growth and vertical mining",
      "text": "FP-growth avoids Apriori’s explicit level-wise candidate generation by compressing transactions into an FP-tree, or frequent-pattern tree. An initial scan determines item supports and removes items below the minimum. The remaining items in each transaction are placed in a consistent order, usually descending support order, and inserted into a tree. Transactions sharing an ordered prefix share tree nodes; node counts record how many transactions follow each prefix. A header structure links nodes bearing the same item.\n\nMining proceeds by conditioning on an item. The paths leading to its nodes form a conditional pattern base, from which a smaller conditional FP-tree can be built. Recursively mining these conditional structures enumerates frequent patterns containing the chosen item. If a conditional tree is a single path, frequent combinations can be generated directly from that path and its counts. The original construction typically uses two full database scans, although the subsequent conditional mining still requires substantial computation and memory access.\n\nThe degree of compression depends on the data. When transactions share many prefixes after ordering, an FP-tree can be much smaller than the original transaction list. When overlap is limited, the tree may remain large. Avoiding candidate generation therefore does not mean avoiding combinatorial growth: an algorithm must still enumerate the patterns it promises to return. A compact input structure can correspond to an exponentially large output.\n\nEclat uses a different representation. Instead of storing each transaction as its items, it stores each itemset together with a transaction-identifier set, or tidset. For itemsets \\(A\\) and \\(B\\),\n\n\\[\\operatorname{tid}(A\\cup B) =\\operatorname{tid}(A)\\cap\\operatorname{tid}(B).\\]\n\nSupport counting becomes set intersection. Eclat commonly explores itemsets using depth-first search, extending prefixes with compatible items and intersecting their tidsets. Sorted identifier lists, bit vectors, and compressed bitmaps offer different time–memory trade-offs. Diffset variants store differences between related transaction sets, which can be compact when extensions remove relatively few transactions.\n\nThese approaches illustrate a broader principle: representation can matter as much as the abstract search strategy. Horizontal transaction scans exploit one access pattern, prefix trees exploit shared structure, and vertical methods exploit fast set operations. Dataset density, transaction length, memory capacity, and the support threshold all influence performance. There is no universal ranking in which one method dominates on every dataset, and comparisons must account for whether implementations produce the same complete pattern collection."
    },
    {
      "heading": "Closed patterns, maximal patterns, and compact rule bases",
      "text": "A frequent itemset collection can contain extensive redundancy. If a set of ten items is frequent, all of its subsets are also frequent, whether or not those subsets add independent information. Compact pattern representations seek to preserve useful information without explicitly storing every frequent set. Two major representations are closed itemsets and maximal frequent itemsets, which preserve different amounts of information.\n\nFor an itemset \\(X\\) with positive support, its closure is the intersection of all transactions containing it:\n\n\\[\\operatorname{cl}(X)=\\bigcap_{j:X\\subseteq T_j}T_j.\\]\n\nThe closure adds every item that always accompanies \\(X\\) in the observed database. An itemset is closed when \\(\\operatorname{cl}(X)=X\\), equivalently when no proper superset has the same support. If every basket containing coffee also contains filters, then coffee and coffee-with-filters have identical support; coffee alone is not closed unless filters are already included in its representation.\n\nThe complete collection of frequent closed itemsets, together with their supports, preserves the support of every frequent itemset. For a frequent \\(X\\), its support equals the maximum support among frequent closed supersets of \\(X\\); its own closure achieves that value. This makes closed patterns a lossless condensed representation for frequent-itemset support information, although retrieving arbitrary supports may require additional indexing or computation.\n\nA maximal frequent itemset, by contrast, has no frequent proper superset. Every maximal frequent itemset is closed, but a closed frequent itemset need not be maximal. Maximal patterns identify the boundary of the frequent region: an itemset is frequent exactly when it is a subset of at least one maximal frequent itemset. Their supports do not, however, determine the supports of all subsets. Consequently, maximal patterns alone are generally insufficient for calculating the confidence of arbitrary rules.\n\nCompact rule bases introduce another distinction. A minimal generator of a closed itemset is a subset having that closure, with no smaller subset sharing it. If \\(G\\) generates \\(C\\), the rule \\(G\\Rightarrow C\\setminus G\\) has confidence one in the observed database. Closure-based methods can use such relationships to represent exact implications and organize approximate rules. Confidence one remains an empirical property; it does not guarantee that future records will obey the implication.\n\nOther redundancy criteria ask whether adding antecedent items meaningfully improves confidence, or whether a rule is merely inherited from a more general relationship. Removing redundancy is not equivalent to keeping only the highest-scoring rules. A compact result may instead preserve coverage, support information, logical structure, or a specified ability to reconstruct omitted rules."
    },
    {
      "heading": "Extensions beyond ordinary co-occurrence",
      "text": "Classical association rules treat items as binary and transactions as unordered. Many applications require a richer description. Quantitative association rules include numerical conditions such as “age between 30 and 39” or “purchase amount above a threshold.” Numerical attributes may be discretized into intervals before mining, or intervals may be searched directly. The choice of boundaries affects both support and apparent rule strength: narrow intervals may isolate strong but poorly supported relationships, while broad intervals may conceal meaningful variation.\n\nMultilevel association rules use an item taxonomy. A transaction containing a particular brand of tea may also be interpreted as containing tea, beverages, and groceries. Higher-level categories usually have greater support than specific products, making a single support threshold awkward across the hierarchy. Mining can use level-dependent thresholds or constraints on which abstraction levels may occur together. Taxonomies also generate trivial rules, such as a product implying its own category, which require separate treatment from genuinely informative associations.\n\nSequential pattern mining preserves ordering across events rather than only set membership. “Customers buy a printer and later buy ink” is different from “printer and ink occur in the same basket.” Sequential methods define support over sequences and may constrain time gaps, event windows, or sequence length. Ordinary association rules cannot recover temporal order once transactions have discarded it. Their directional notation alone does not imply that the antecedent happened earlier.\n\nHigh-utility itemset mining incorporates quantities and item-specific values such as unit profit. A low-frequency combination can then be important because of its total utility. Unlike support, ordinary positive transaction utility is not generally anti-monotone under itemset extension: adding an item can increase utility. Algorithms therefore use suitable upper bounds, such as transaction-weighted utility under appropriate assumptions, to prune safely. Simply substituting profit for support in Apriori does not preserve its correctness.\n\nNegative rules describe absence, such as \\(X\\Rightarrow\\neg Y\\). Their meaning depends on whether absence is genuinely observed rather than merely unrecorded. A missing diagnosis code, for example, need not establish that a condition is absent. Negative patterns can also proliferate because most transactions omit most possible items.\n\nClass association rules restrict consequents to class labels. In associative classification, such rules are selected and organized into a predictive model, placing the task within supervised learning. This requires procedures for resolving competing rules and evaluating predictions on unseen cases. More generally, constraint-based mining can restrict antecedent length, permitted items, taxonomic combinations, or consequent attributes, making the output better aligned with a defined analytical question."
    },
    {
      "heading": "Applications and operational interpretation",
      "text": "Market-basket analysis remains the canonical application. Rules can describe products frequently purchased together and inform investigations of assortment, promotions, or cross-selling. Their operational meaning depends on the underlying process: a bundle sold under one package identifier, a mandatory accessory, or a temporary discount can all produce strong co-occurrence for different reasons. Transaction records establish what was bought together under existing conditions; they do not by themselves establish how changing a display or price would alter purchases.\n\nIn a recommender system, the items already selected by a user can match rule antecedents, while consequents provide candidate recommendations. Rules are interpretable and can work without explicit user profiles. However, association strength is not the same as recommendation relevance. An item may already have been purchased, be unavailable, or be inappropriate for repeat recommendation. A practical system must combine rule scores with eligibility constraints, stock information, and an evaluation objective such as ranking quality or conversion.\n\nWeb and application usage analysis treats page visits, actions, or session attributes as items. An association might connect use of a particular feature with completion of a workflow. Session construction is crucial because automated traffic, shared accounts, and inactivity cutoffs can alter which events are grouped. When order is important—for example, distinguishing an error preceding abandonment from an error encountered after a different task—sequence-based analysis is more appropriate than ordinary co-occurrence mining.\n\nIn manufacturing and maintenance, transactions can represent inspection records, equipment configurations, or time windows of sensor states. Rules may reveal combinations of settings associated with defects or groups of alarms that recur together. Such findings support investigation, but alarms often share a common underlying cause or reflect propagation through the same system. In cybersecurity, analogous methods can identify recurring event combinations or help characterize unusual deviations; the reliability of the result depends heavily on event normalization and background activity.\n\nScientific datasets can also be represented transactionally. Molecular measurements may be converted into categorical states, and observational records may be searched for recurring combinations. Here, discretization, batch effects, and the number of comparisons become central concerns. A discovered rule is a candidate descriptive relationship rather than an independently established mechanism.\n\nAcross these applications, practical value is external to the basic support–confidence problem. A rule can be statistically strong but already obvious, impossible to act upon, or economically irrelevant. Conversely, a modestly frequent relationship can matter when it exposes a costly failure mode. Evaluation therefore distinguishes computational validity, empirical reproducibility, novelty to the domain, and the consequences of using a rule in an operational system."
    },
    {
      "heading": "Statistical validation and limits of interpretation",
      "text": "Association mining is an adaptive search over many possible relationships. Even when items are independent in the generating population, some combinations will look strongly associated in a finite sample. Examining millions of candidates and reporting only the most impressive ones creates a multiple-comparison problem. Minimum support limits the search and removes extremely small counts, but neither support nor confidence thresholds alone provide a statistical guarantee against false discoveries.\n\nFor a fixed pair of occurrence events, their relationship can be represented by a \\(2\\times2\\) contingency table: both occur, only the antecedent occurs, only the consequent occurs, or neither occurs. Statistical tests can assess departure from an independence model. However, the validity of the sampling assumptions matters. Repeated transactions from the same customer, overlapping sensor windows, or correlated observations within a site do not behave like independent draws. Enormous transaction counts can also make tiny, operationally unimportant departures statistically significant.\n\nMultiple-testing procedures control specified error rates across collections of hypotheses. Their use in pattern mining requires care because the hypothesis family may be enormous and the search itself depends on observed supports. Methods that exploit testability bounds, resampling, or carefully specified candidate families can reduce the burden. A significance value computed after an unrestricted search should not be interpreted as if the rule had been selected independently of the data.\n\nIndependent validation addresses a different but complementary issue. Rules can be discovered in training data, selected using a validation set, and finally assessed on a held-out test set. This reduces overfitting to accidental combinations. For temporal applications, an earlier discovery period and a later evaluation period often better represent deployment than a random split. Customer- or device-level separation may be necessary to prevent data leakage through repeated entities.\n\nAssociation also does not establish causation. A promotion can cause two products to be purchased together; seasonality can increase both; and selection into a dataset can create an association that is absent in the wider population. Conditional relationships may differ across subgroups or reverse when aggregated, an effect associated with Simpson’s paradox. Conditioning on more attributes can reveal structure, but it also reduces sample sizes and expands the search space.\n\nFinally, apparent rule stability can disappear under distribution shift. Product catalogs, user behavior, logging systems, and institutional practices change. Reproducibility therefore involves specifying the population, observation period, transaction definition, preprocessing, thresholds, and evaluation procedure—not merely reporting a rule and its confidence."
    },
    {
      "heading": "Scalability, data governance, and reproducible implementation",
      "text": "The search space imposes a fundamental limitation. With \\(m\\) items, there are \\(2^m-1\\) nonempty itemsets. Even an ideal implementation cannot explicitly return an exponentially large result in sub-output time. This links association mining to computational complexity and makes output control a first-class design issue. Minimum support, maximum pattern length, restricted consequents, closed representations, and top-\\(k\\) objectives address different versions of that issue. A maximum length restriction makes a result incomplete for unrestricted mining but can be appropriate for a precisely defined task.\n\nLarge-scale implementations use partitioning and distributed computing to share counting and search work. In a simple distributed counting scheme, workers count candidates within local partitions and aggregate counts globally. Any itemset meeting a global relative support threshold must meet that same relative threshold in at least one partition, which supports partition-based candidate discovery. Local candidate unions can nevertheless be large, and global verification still incurs data access or communication.\n\nPattern-growth approaches can partition work by suffix or prefix, but task sizes may be highly unequal. A very common item can induce a much larger conditional problem than a rare one. Systems built around MapReduce or Apache Spark must therefore manage communication, data replication, memory use, and skew, not just arithmetic. Hardware acceleration of set intersections can help vertical methods, but faster counting cannot eliminate the cost of enumerating a huge output.\n\nData quality is equally important. Duplicate imports inflate support; inconsistent item identifiers split one concept into several; changing catalogs create artificial absences; and omitted transactions change denominators. In transaction mining, “transaction” means an analytical record and need not coincide with a database transaction in the sense of database concurrency and recovery. Reproducible implementations preserve the distinction and document how source records become analytical baskets.\n\nAssociation results can disclose sensitive information even when direct identifiers are absent. Rare combinations may correspond to small groups, while rules can reveal relationships involving sensitive attributes. Data privacy therefore concerns both input access and released patterns. Differential privacy offers a formal framework for limiting how much an output depends on one protected unit. Private frequent-pattern mining must manage noisy counts, candidate selection, and the size of the search space. Whether the protected unit is one transaction or an individual’s entire history materially changes the guarantee.\n\nFor reproducibility, implementations record absolute and relative thresholds, item mappings, missing-value treatment, transaction boundaries, software versions, and rule-filtering criteria. Reporting supporting counts alongside derived measures makes rounding and denominator choices visible. Where records evolve over time, data lineage connects a reported association to the source data and transformations that produced it."
    }
  ],
  "links": [
    "data-mining",
    "machine-learning",
    "unsupervised-learning",
    "relational-database",
    "probability",
    "ibm",
    "hash-table",
    "depth-first-search",
    "supervised-learning",
    "recommender-system",
    "cybersecurity",
    "multiple-testing",
    "training-data",
    "validation-set",
    "test-set",
    "overfitting",
    "data-leakage",
    "causation",
    "distribution-shift",
    "computational-complexity",
    "distributed-computing",
    "mapreduce",
    "apache-spark",
    "data-quality",
    "database-transaction",
    "data-privacy",
    "differential-privacy",
    "data-lineage"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}