{ "profile": "wrs-published-research-view-1", "edition": "ratings-e039106f30dec177", "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS", "entry": { "id": "kanter-veeramachaneni-deep-feature-synthesis-data-science-machine-2015", "title": "Deep Feature Synthesis: constructing predictors from related tables", "summary": "A prediction problem often begins with several connected tables: customers, purchases and individual items, for example. Useful inputs may include a customer’s purchase count or the average price of their items, rather than just one table column. Kanter and Veeramachaneni describe Deep Feature Synthesis, which composes transformations and aggregations along table relationships. Their Data Science Machine combines these generated features with selection, preprocessing, model fitting and tuning. The result is a reusable way to automate part of feature construction; “deep” refers to composed features, not necessarily a deep neural network.", "boundary": "The relational schema, prediction target and time-dependent filters still need setup. The reported comparison covers three competitions, including snapshots of competitions still running; outperforming participating teams is not replacing that number of scientists. AUC comparisons are not ordinary accuracy percentages. Feature growth, computation and proper separation of training and validation data remain important. The paper reports manual troubleshooting of a submission mismatch; no universal autonomous data-science or leakage-free evaluation guarantee is implied.", "area": "Automated feature engineering", "topic": "Artificial intelligence", "year": 2015, "lane": "HUMAN", "result": "ALGORITHM", "package_id": "kanter-veeramachaneni-deep-feature-synthesis-data-science-machine-2015", "target": { "record_id": "urn:uuid:7343bf68-bfc3-4eb6-a8d2-82bcf076b777", "version": 1, "content_hash": "sha256:9094d76fb6d5d13a3b832103fe03890c996338f2467cfe57910bfb4ba33966df" }, "humans": [ "James Max Kanter", "Kalyan Veeramachaneni" ], "authors_complete": true, "human_role_note": "Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred.", "agents": [], "related": [], "subresults": [], "scientific_assessment": { "source_review": "SOURCE_CURATED", "scientific_review": "NOT_REVIEWED_BY_WRS", "formal_replay": "NOT_RUN", "provider_attestation": "NONE" }, "sources": [ { "id": "cur20261017-mit-deep-feature-synthesis-2015", "title": "Deep Feature Synthesis: Towards Automating Data Science Endeavors", "url": "https://groups.csail.mit.edu/EVO-DesignOpt/groupWebSite/uploads/Site/DSAA_DSM_2015.pdf", "date": "2015" } ], "geography": { "country_codes": [ "US" ], "location_basis": "Original CSAIL MIT byline explicitly Cambridge, MA. Work affiliation, not nationality or ownership.", "source_url": "https://groups.csail.mit.edu/EVO-DesignOpt/groupWebSite/uploads/Site/DSAA_DSM_2015.pdf", "target": { "record_id": "urn:uuid:7343bf68-bfc3-4eb6-a8d2-82bcf076b777", "version": 1, "content_hash": "sha256:9094d76fb6d5d13a3b832103fe03890c996338f2467cfe57910bfb4ba33966df" }, "basis": "RESEARCH_LOCATION_OR_PAPER_AFFILIATION", "reviewed_on": "2026-10-23", "nationality_inferred": false, "sources": [ { "title": "Deep Feature Synthesis: Towards Automating Data Science Endeavors", "url": "https://groups.csail.mit.edu/EVO-DesignOpt/groupWebSite/uploads/Site/DSAA_DSM_2015.pdf" } ], "coverage_status": "COMPLETE", "country_labels": { "US": "United States" } }, "points": "100000", "published_grade": { "id": "kanter-veeramachaneni-deep-feature-synthesis-data-science-machine-2015", "entry_id": "kanter-veeramachaneni-deep-feature-synthesis-data-science-machine-2015", "band": "A", "modifier": null, "target": { "record_id": "urn:uuid:7343bf68-bfc3-4eb6-a8d2-82bcf076b777", "version": 1, "content_hash": "sha256:9094d76fb6d5d13a3b832103fe03890c996338f2467cfe57910bfb4ba33966df" }, "what_changed": "A prediction problem often begins with several connected tables: customers, purchases and individual items, for example. Useful inputs may include a customer’s purchase count or the average price of their items, rather than just one table column. Kanter and Veeramachaneni describe Deep Feature Synthesis, which composes transformations and aggregations along table relationships. Their Data Science Machine combines these generated features with selection, preprocessing, model fitting and tuning. The result is a reusable way to automate part of feature construction; “deep” refers to composed features, not necessarily a deep neural network.", "scope": "A prediction problem often begins with several connected tables: customers, purchases and individual items, for example. Useful inputs may include a customer’s purchase count or the average price of their items, rather than just one table column. Kanter and Veeramachaneni describe Deep Feature Synthesis, which composes transformations and aggregations along table relationships. Their Data Science Machine combines these generated features with selection, preprocessing, model fitting and tuning. The result is a reusable way to automate part of feature construction; “deep” refers to composed features, not necessarily a deep neural network. The relational schema, prediction target and time-dependent filters still need setup. The reported comparison covers three competitions, including snapshots of competitions still running; outperforming participating teams is not replacing that number of scientists. AUC comparisons are not ordinary accuracy percentages. Feature growth, computation and proper separation of training and validation data remain important. The paper reports manual troubleshooting of a submission mismatch; no universal autonomous data-science or leakage-free evaluation guarantee is implied.", "status": "PROVISIONAL_EDITORIAL", "scientific_correctness_certified": false }, "assessment_disposition": { "status": "ASSESSED", "label": "Provisionally assessed", "reason": "A current grade and separate reference score identify this exact contribution.", "target": { "record_id": "urn:uuid:7343bf68-bfc3-4eb6-a8d2-82bcf076b777", "version": 1, "content_hash": "sha256:9094d76fb6d5d13a3b832103fe03890c996338f2467cfe57910bfb4ba33966df" }, "edition": "ratings-e039106f30dec177" } }, "version": { "kind": "PUBLIC_RECORD_VIEW", "previous": null, "published_at": "2026-10-23T00:00:00Z", "publisher": { "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33", "role": "WRS_EDITORIAL_PUBLICATION" }, "record": { "contributors": [ { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "James Max Kanter", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Deep Feature Synthesis: Towards Automating Data Science Endeavors", "url": "https://groups.csail.mit.edu/EVO-DesignOpt/groupWebSite/uploads/Site/DSAA_DSM_2015.pdf" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Kalyan Veeramachaneni", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Deep Feature Synthesis: Towards Automating Data Science Endeavors", "url": "https://groups.csail.mit.edu/EVO-DesignOpt/groupWebSite/uploads/Site/DSAA_DSM_2015.pdf" } } ], "evidence": [ { "description": "Bibliographic/stated-result source; not a WRS proof verification", "id": "source-0", "kind": "source_reference", "source": { "citation": "Deep Feature Synthesis: Towards Automating Data Science Endeavors", "url": "https://groups.csail.mit.edu/EVO-DesignOpt/groupWebSite/uploads/Site/DSAA_DSM_2015.pdf" } } ], "origin": "HUMAN_LEGACY", "record_profile": "wrs-local-research-notes-0.1", "record_type": "research_note", "rrp_version": "0.1", "scope": { "description": "The relational schema, prediction target and time-dependent filters still need setup. The reported comparison covers three competitions, including snapshots of competitions still running; outperforming participating teams is not replacing that number of scientists. AUC comparisons are not ordinary accuracy percentages. Feature growth, computation and proper separation of training and validation data remain important. The paper reports manual troubleshooting of a submission mismatch; no universal autonomous data-science or leakage-free evaluation guarantee is implied.", "domain": "Automated feature engineering" }, "sources": [ { "citation": "Deep Feature Synthesis: Towards Automating Data Science Endeavors", "url": "https://groups.csail.mit.edu/EVO-DesignOpt/groupWebSite/uploads/Site/DSAA_DSM_2015.pdf" } ], "statement": "# Deep Feature Synthesis: constructing predictors from related tables\n\n## The useful number may not be a column\n\nSuppose a shop wants to predict whether a customer will return next month. Its customer table has names and registration dates, while purchases live in a different table. A useful clue might be the customer’s number of purchases or average basket value. Neither has to exist as a ready-made column. Deep Feature Synthesis explores how to construct such inputs by following relationships between tables and combining simple operations.\n\n## Follow the relationship, then summarize\n\nA feature is a quantity supplied to a prediction model. An aggregation summarizes several related rows, for example by counting purchases or averaging prices. A transformation changes an existing quantity, such as extracting a month from a date. The method can compose these operations through a relational schema, which describes how tables connect. “Deep” refers to that composition; it does not mean the system must be a deep neural network.\n\n## Build one feature by hand first\n\nUse an invented classroom customer with three earlier purchases of ten, twenty and thirty pounds. The purchase count is three. The average purchase value is sixty divided by three, or twenty pounds. A second customer with one sixty-pound purchase has the same total spending but a different count and average. These features capture different behavior. The example is not a commercial result from the paper and says nothing by itself about which customer will return.\n\n## A prediction must respect the clock\n\nIf you predict a decision made on the first day of a month, a purchase made later that month cannot be an input. Using it leaks information from the future. Schema design, cutoff dates and separation of training from validation therefore remain important even when feature generation is automated. More generated features also need storage and computation. The original integrated system includes feature selection and tuning; it does not remove every human decision or data-quality problem.\n\n## Inspect a small table before a big claim\n\nAsk your agent to create two related toy tables and list three features using only information available before a chosen date. Add one deliberately leaking feature and explain why to reject it. Require intermediate rows and arithmetic, not just final model accuracy. The paper’s competition comparisons demonstrate a bounded use of the pipeline; outperforming a number of competition teams does not establish replacement of that many scientists or autonomous discovery of business problems.\n\n## Scientific scope\n\nThe relational schema, prediction target and time-dependent filters still need setup. The reported comparison covers three competitions, including snapshots of competitions still running; outperforming participating teams is not replacing that number of scientists. AUC comparisons are not ordinary accuracy percentages. Feature growth, computation and proper separation of training and validation data remain important. The paper reports manual troubleshooting of a submission mismatch; no universal autonomous data-science or leakage-free evaluation guarantee is implied.\n\n## Sources for this explanation\n\n- [Deep Feature Synthesis: Towards Automating Data Science Endeavors](https://groups.csail.mit.edu/EVO-DesignOpt/groupWebSite/uploads/Site/DSAA_DSM_2015.pdf): Original 2015 paper: relational feature construction, integrated pipeline, competition setup and the reported local-versus-submitted prediction discrepancy.\n", "summary": "A prediction problem often begins with several connected tables: customers, purchases and individual items, for example. Useful inputs may include a customer’s purchase count or the average price of their items, rather than just one table column. Kanter and Veeramachaneni describe Deep Feature Synthesis, which composes transformations and aggregations along table relationships. Their Data Science Machine combines these generated features with selection, preprocessing, model fitting and tuning. The result is a reusable way to automate part of feature construction; “deep” refers to composed features, not necessarily a deep neural network.", "title": "Deep Feature Synthesis: constructing predictors from related tables" }, "record_id": "urn:uuid:7343bf68-bfc3-4eb6-a8d2-82bcf076b777", "version": 1, "source_ref": { "record_id": "urn:uuid:7343bf68-bfc3-4eb6-a8d2-82bcf076b777", "version": 1, "content_hash": "sha256:9094d76fb6d5d13a3b832103fe03890c996338f2467cfe57910bfb4ba33966df" }, "export_view": "wrs-public-editorial-view-1", "public_view_hash": "sha256:457d090191df9be45075cf5a0a3a68d4a741763bdcb8dfa21f1675c8d6d70d3f" }, "target": { "record_id": "urn:uuid:7343bf68-bfc3-4eb6-a8d2-82bcf076b777", "version": 1, "content_hash": "sha256:9094d76fb6d5d13a3b832103fe03890c996338f2467cfe57910bfb4ba33966df" }, "published_grade": { "profile": "wrs-published-grade-1", "edition": "2026-10-28.1", "id": "kanter-veeramachaneni-deep-feature-synthesis-data-science-machine-2015", "entry_id": "kanter-veeramachaneni-deep-feature-synthesis-data-science-machine-2015", "revision": 1, "title": "Deep Feature Synthesis: constructing predictors from related tables", "context": "Automated feature engineering", "track": "Reusable scientific method, experimental result or historical framework", "band": "A", "modifier": null, "status": "PROVISIONAL_EDITORIAL", "assessed_on": "2026-10-23", "evidence_cutoff": "2026-10-23", "achievement_state": "SUPPORTED_SCOPED_RESULT", "what_changed": "A prediction problem often begins with several connected tables: customers, purchases and individual items, for example. Useful inputs may include a customer’s purchase count or the average price of their items, rather than just one table column. Kanter and Veeramachaneni describe Deep Feature Synthesis, which composes transformations and aggregations along table relationships. Their Data Science Machine combines these generated features with selection, preprocessing, model fitting and tuning. The result is a reusable way to automate part of feature construction; “deep” refers to composed features, not necessarily a deep neural network.", "scope": "A prediction problem often begins with several connected tables: customers, purchases and individual items, for example. Useful inputs may include a customer’s purchase count or the average price of their items, rather than just one table column. Kanter and Veeramachaneni describe Deep Feature Synthesis, which composes transformations and aggregations along table relationships. Their Data Science Machine combines these generated features with selection, preprocessing, model fitting and tuning. The result is a reusable way to automate part of feature construction; “deep” refers to composed features, not necessarily a deep neural network. The relational schema, prediction target and time-dependent filters still need setup. The reported comparison covers three competitions, including snapshots of competitions still running; outperforming participating teams is not replacing that number of scientists. AUC comparisons are not ordinary accuracy percentages. Feature growth, computation and proper separation of training and validation data remain important. The paper reports manual troubleshooting of a submission mismatch; no universal autonomous data-science or leakage-free evaluation guarantee is implied.", "evidence_summary": "The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. The relational schema, prediction target and time-dependent filters still need setup. The reported comparison covers three competitions, including snapshots of competitions still running; outperforming participating teams is not replacing that number of scientists. AUC comparisons are not ordinary accuracy percentages. Feature growth, computation and proper separation of training and validation data remain important. The paper reports manual troubleshooting of a submission mismatch; no universal autonomous data-science or leakage-free evaluation guarantee is implied.", "references": [ { "title": "Deep Feature Synthesis: Towards Automating Data Science Endeavors", "url": "https://groups.csail.mit.edu/EVO-DesignOpt/groupWebSite/uploads/Site/DSAA_DSM_2015.pdf", "locator": "Original scoped contribution, statement/method and evidence limits; source-specific attribution retained." } ], "roles": [ { "who": "The named authors in the research record", "credit": "SCIENCE", "did": "Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred." } ], "target": { "record_id": "urn:uuid:7343bf68-bfc3-4eb6-a8d2-82bcf076b777", "version": 1, "content_hash": "sha256:9094d76fb6d5d13a3b832103fe03890c996338f2467cfe57910bfb4ba33966df" }, "scientific_correctness_certified": false, "independent_expert_calibration": false, "method": "PROPRIETARY", "points": "100000" }, "canonical_agent_submission": false, "scientific_verification_awarded": false, "reader_guide": { "id": "kanter-veeramachaneni-deep-feature-synthesis-data-science-machine-2015", "target": { "record_id": "urn:uuid:7343bf68-bfc3-4eb6-a8d2-82bcf076b777", "version": 1, "content_hash": "sha256:9094d76fb6d5d13a3b832103fe03890c996338f2467cfe57910bfb4ba33966df" }, "sections": [ { "title": "The useful number may not be a column", "text": "Suppose a shop wants to predict whether a customer will return next month. Its customer table has names and registration dates, while purchases live in a different table. A useful clue might be the customer’s number of purchases or average basket value. Neither has to exist as a ready-made column. Deep Feature Synthesis explores how to construct such inputs by following relationships between tables and combining simple operations." }, { "title": "Follow the relationship, then summarize", "text": "A feature is a quantity supplied to a prediction model. An aggregation summarizes several related rows, for example by counting purchases or averaging prices. A transformation changes an existing quantity, such as extracting a month from a date. The method can compose these operations through a relational schema, which describes how tables connect. “Deep” refers to that composition; it does not mean the system must be a deep neural network." }, { "title": "Build one feature by hand first", "text": "Use an invented classroom customer with three earlier purchases of ten, twenty and thirty pounds. The purchase count is three. The average purchase value is sixty divided by three, or twenty pounds. A second customer with one sixty-pound purchase has the same total spending but a different count and average. These features capture different behavior. The example is not a commercial result from the paper and says nothing by itself about which customer will return." }, { "title": "A prediction must respect the clock", "text": "If you predict a decision made on the first day of a month, a purchase made later that month cannot be an input. Using it leaks information from the future. Schema design, cutoff dates and separation of training from validation therefore remain important even when feature generation is automated. More generated features also need storage and computation. The original integrated system includes feature selection and tuning; it does not remove every human decision or data-quality problem." }, { "title": "Inspect a small table before a big claim", "text": "Ask your agent to create two related toy tables and list three features using only information available before a chosen date. Add one deliberately leaking feature and explain why to reject it. Require intermediate rows and arithmetic, not just final model accuracy. The paper’s competition comparisons demonstrate a bounded use of the pipeline; outperforming a number of competition teams does not establish replacement of that many scientists or autonomous discovery of business problems." } ], "references": [ { "title": "Deep Feature Synthesis: Towards Automating Data Science Endeavors", "url": "https://groups.csail.mit.edu/EVO-DesignOpt/groupWebSite/uploads/Site/DSAA_DSM_2015.pdf", "locator": "Original 2015 paper: relational feature construction, integrated pipeline, competition setup and the reported local-versus-submitted prediction discrepancy." } ] }, "reading_status": "EXPLANATORY_GUIDE" }