{
  "id": "crisp-dm",
  "title": "CRISP-DM",
  "lang": "en",
  "source_lang": "en",
  "summary": "CRISP-DM is an industry-neutral framework for planning and carrying out data-mining projects through six iterative phases, from business understanding to deployment.",
  "category": "computer-science",
  "subcategory": "artificial-intelligence",
  "type": "concept",
  "aliases": [
    "Cross-Industry Standard Process for Data Mining",
    "CRISP-DM 1.0"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/crisp-dm/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/crisp-dm/",
      "json": "https://aiwiki.page/zh/crisp-dm.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "CRISP-DM (Cross-Industry Standard Process for Data Mining) is a process framework for organizing data-mining projects. It describes six interconnected phases: business understanding, data understanding, data preparation, modeling, evaluation, and deployment. Rather than prescribing a particular algorithm, software product, or industry-specific solution, it provides a common structure for defining objectives, examining data, developing analytical results, and putting those results into use. Its central principle is that successful analysis requires more than a technically effective model: the work must address an understood practical problem and produce results that can be evaluated and applied."
    },
    {
      "heading": "Origins and scope",
      "text": "CRISP-DM emerged from an initiative begun in 1996 involving Daimler-Benz, SPSS, NCR, and the Dutch insurer OHRA. The CRISP-DM 1.0 reference model and user guide were published in 2000. The framework combined experience from industrial data-mining projects into a methodology intended to work across organizations, application domains, and software tools.\n\nThe word standard in its name describes its intended role as a shared process model. It does not mean that following CRISP-DM constitutes certification under a formal international standards scheme.\n\nAlthough developed for data mining, the framework can also organize projects involving machine learning, statistical analysis, and related analytical methods. Its scope is broader than model training: it includes problem definition, feasibility assessment, communication, operational use, and maintenance planning. Conversely, it does not supply a complete technical architecture for data platforms or production systems."
    },
    {
      "heading": "Structure of the framework",
      "text": "CRISP-DM distinguishes several levels of description:\n\n- Phases identify the major kinds of work in a project.\n- Generic tasks describe activities that apply across many projects.\n- Specialized tasks adapt those activities to particular technical or organizational circumstances.\n- Process instances record how the work was actually performed in an individual project.\n\nThe reference model identifies tasks and their expected outputs, while the user guide provides more detailed guidance on carrying them out. Outputs can include business objectives, data-quality reports, modeling assumptions, evaluation findings, and deployment plans.\n\nThe six phases are not a rigid sequence. Data preparation may expose a problem that requires renewed data understanding; modeling may reveal that additional variables are needed; evaluation may show that the original business objective was poorly specified. The framework also treats the project lifecycle as cyclical: experience from deployment can lead to a new analytical project."
    },
    {
      "heading": "The six phases",
      "text": "1. Business understanding\n\nThe first phase establishes what the organization or other project sponsor wants to achieve and translates that objective into an analytical problem.\n\nIts principal tasks are to:\n\n- Determine business objectives.\n- Assess the situation, including available resources, constraints, risks, and assumptions.\n- Determine data-mining goals.\n- Produce a project plan.\n\nA business objective and a modeling objective are related but distinct. For example, reducing customer attrition is a business objective; estimating which customers are likely to leave within a specified period is a possible data-mining objective. Even an accurate estimate does not establish that a particular retention intervention will work.\n\nSuccess criteria therefore need to cover both analytical performance and practical usefulness. Constraints may include delivery deadlines, the availability of staff and data, operational costs, and restrictions on data use. This phase can also establish that a project is infeasible or that a simpler non-modeling approach is sufficient.\n\n2. Data understanding\n\nData understanding begins with collecting or accessing an initial dataset. Its tasks include describing the data, exploring it, and verifying data quality.\n\nTypical questions include:\n\n- What does each field represent, and how was it recorded?\n- Which populations, events, or time periods are represented?\n- Are values missing, inconsistent, duplicated, or implausible?\n- Is the proposed target variable measured reliably?\n- Does the available information correspond to what would be available when the result is used?\n\nExploration may use descriptive statistics, data visualization, and comparisons across groups or time periods. Knowledge of metadata and data lineage helps establish whether apparently useful variables have an appropriate interpretation.\n\nThis phase is not merely a preliminary inspection. Discoveries about coverage, measurement, or collection procedures can change the project’s scope and the meaning of its eventual conclusions.\n\n3. Data preparation\n\nData preparation constructs the dataset or datasets used in modeling. CRISP-DM groups this work into selecting, cleaning, constructing, integrating, and formatting data.\n\nActivities can include joining records from multiple sources, resolving inconsistent values, handling missing observations, and carrying out feature engineering. Depending on the chosen method, preparation may also involve feature scaling, encoding categorical variables, or feature selection.\n\nThe boundary between preparation and modeling is iterative. A method may require a different representation of the data, while preparation may reveal that the proposed method is unsuitable.\n\nAn important methodological concern is data leakage. For example, estimating preprocessing parameters from the full dataset before separating evaluation data can make measured performance misleading. Transformations learned from data generally need to be fitted on the appropriate training data and then applied consistently to held-out observations. Temporal and grouped datasets may require separation rules that reflect their underlying structure.\n\n4. Modeling\n\nThe modeling phase selects techniques, defines a test design, builds models, and assesses their technical properties.\n\nPossible techniques include supervised learning, unsupervised learning, and other statistical or data-mining methods. CRISP-DM does not favor a particular model family. Selection depends on the analytical goal, data characteristics, interpretability requirements, and practical constraints.\n\nThe test design establishes how performance will be assessed. Model development can involve comparing candidates and adjusting hyperparameters, often with a validation set or resampling procedure. A separate test set can support a final assessment that has not been repeatedly used to guide model selection.\n\nTechnical assessment examines performance, assumptions, stability, and problems such as overfitting. This is narrower than the next phase: a model can perform well under a statistical criterion without being suitable for the intended business use.\n\n5. Evaluation\n\nEvaluation considers whether the analytical results meet the business objectives and whether important issues have been overlooked.\n\nThe phase includes:\n\n- Evaluating results against business success criteria.\n- Reviewing the process used to obtain them.\n- Determining the next steps.\n\nThis distinction between model assessment and project evaluation is central to CRISP-DM. High predictive accuracy may be insufficient if a model is too costly to operate, depends on information unavailable at decision time, or produces outputs that do not support an effective action.\n\nEvaluation may lead to deployment, further modeling, additional data collection, revision of the objectives, or termination of the project. It is not simply a final approval step after technical work has finished.\n\n6. Deployment\n\nDeployment puts the project’s results into practical use. It can be as simple as delivering a report or as complex as integrating a model into an operational decision system.\n\nThe framework calls for deployment planning, monitoring and maintenance planning, a final report, and a project review. Consequently, deployment is not synonymous with exposing a prediction service through an application programming interface. A dashboard, revised procedure, or documented analytical finding may also be a deployment outcome.\n\nFor operational models, monitoring can include input quality, service performance, and changes in predictive behavior. Distribution shift can reduce the usefulness of a model after deployment, making reassessment or retraining necessary. The original framework recognizes maintenance needs but does not specify a detailed automated production infrastructure."
    },
    {
      "heading": "Iteration and project documentation",
      "text": "CRISP-DM is sometimes represented as six boxes around a lifecycle diagram. That representation can obscure its flexibility: the work is expected to move backward and forward between phases.\n\nFor example, a model may appear effective until evaluation reveals that a strong predictor is recorded only after the event being predicted. Addressing the problem may require returning to data understanding, rebuilding the prepared dataset, and repeating modeling. In another project, deployment constraints may require replacing a complex model with a simpler one.\n\nDocumentation connects these iterations. Recording objectives, data definitions, preparation decisions, assumptions, and evaluation outcomes makes it possible to explain why a result was accepted or rejected. Version control and other reproducibility tools can support this work, although CRISP-DM does not prescribe specific tools."
    },
    {
      "heading": "Applications and limitations",
      "text": "CRISP-DM provides a shared vocabulary for analysts, domain specialists, managers, and technical teams. It is particularly useful when the main challenge includes translating an organizational problem into an analytical task rather than merely implementing a known algorithm. Its separation of business objectives from technical metrics also makes explicit why model quality alone does not establish project success.\n\nThe framework nevertheless requires adaptation:\n\n- It is high-level. It does not determine the appropriate sampling design, evaluation metric, causal method, or algorithm for a particular problem.\n- It is not a complete engineering methodology. Production implementation often requires additional software engineering and data engineering practices, including testing, dependency management, release processes, and operational ownership.\n- Its original treatment of governance is limited. Contemporary projects may need more explicit procedures for data privacy, security, accountability, and the assessment of unequal effects across groups.\n- It does not by itself ensure valid inference. Reliable prediction does not establish causation, and using the framework cannot compensate for inadequate measurements or an unsuitable study design.\n- Its lifecycle abstraction needs interpretation for continuous systems. Streaming applications and frequently updated models require recurring activities rather than a single passage through six phases.\n\nCRISP-DM is therefore a framework for structuring analytical work, not a guarantee of successful analysis or a substitute for domain knowledge and methodological judgment."
    }
  ],
  "links": [
    "data-mining",
    "algorithm",
    "machine-learning",
    "statistics",
    "data-quality",
    "data-visualization",
    "metadata",
    "data-lineage",
    "feature-engineering",
    "feature-scaling",
    "feature-selection",
    "data-leakage",
    "training-data",
    "supervised-learning",
    "unsupervised-learning",
    "hyperparameter",
    "validation-set",
    "test-set",
    "overfitting",
    "application-programming-interface",
    "distribution-shift",
    "version-control",
    "software-engineering",
    "data-engineering",
    "data-privacy",
    "causation"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}