{"id":"390b7827-2ddc-427c-a721-8aebd6d63755","arxiv_id":"2501.10555","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey of feature selection and feature generation for tabular data, categorizing traditional, RL-based, and generative AI methods into a unified taxonomy.","lead":"This paper is a survey of methods that improve tabular data by selecting and generating features, covering traditional, reinforcement learning, and generative AI approaches. It gives researchers and practitioners a structured map of the field with a taxonomy, comparisons, and future directions.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Citation-reference mismatches and a missing citation make the survey's core claim of being a reliable, comprehensive reference unverifiable as written.","rationale":"The reader's weakest_assumption is that the survey's descriptions and citations of prior methods are accurate and correctly attributed. This is indeed the load-bearing condition for a survey whose stated contribution is to integrate feature selection and generation into one reliable framework. The manuscript itself provides direct evidence that this condition is violated: in-text citations disagree with Figure 2, the reference list contains duplicate entries, and at least one citation is a literal placeholder. These are not matters of interpretation or outside-consensus disagreement; they are objective, internally checkable failures. The reader's CONDITIONAL verdict is appropriate because the high-level organization and comparative tables may still be useful after corrections, but the paper as submitted cannot serve as a trustworthy entry point to the literature. I agree with the reader's identification of the weakest assumption and see no need to move the verdict.","tokens_in":34602,"tokens_out":2803,"duration_ms":28652,"concrete_test":"Extract every bracketed citation from the text and Figure 2, map each to the reference list, and compare the cited reference's title and venue to the claim in the surrounding sentence; then manually verify a random sample of 50 such mappings. If more than 10% of the sampled claims cite a reference whose title or topic does not match the described method, or if any claimed method has no resolvable citation (e.g., mRMR), the survey's reliability claim fails and the paper must be revised before acceptance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support its central claim of being a comprehensive, unified survey, the paper must accurately map each described method to a verifiable source. This condition fails visibly and pervasively. In §3.1.1, the chi-square test is cited as [38,99] while Figure 2 attributes it to [45,46]; ANOVA is cited as [77,132,155] but Figure 2 says [47,48,49]; Pearson correlation is [193] in the text versus [50] in Figure 2; mutual information is [5,193] versus [50,51]; and mRMR carries a literal '[? ]' placeholder. The reference list itself contains duplicates: [32] and [33] are the same Fan et al. paper, [54] and [55] are the same Huang et al. paper, and [91] and [92] are the same Liu et al. paper. Figure 2's citation blocks also conflict with the text, for example the 'Advanced Statistical Summaries & GCNs State Representation' item cites [39] while the corresponding text in §5.1.1 attributes the GCN-based multi-agent work to [89,90]. Since a survey's utility is precisely that readers can locate and trust the cited methods, these mismatches mean the central claim of comprehensiveness is currently unverifiable. The concern is not that the taxonomy is wrong, but that the evidence base for it cannot be checked.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper surveys tabular data-centric AI, focusing on feature selection and feature generation. It organizes feature selection into single-view (filter, wrapper, embedded, hybrid) and multi-view (supervised, semi-supervised, unsupervised) methods, and feature generation into human-driven and automated techniques. It then reviews advanced methods based on reinforcement learning and generative AI, provides an MDP formulation for RL-based feature engineering, an encoder-decoder-evaluator formulation for generative approaches, a comparative analysis with practical guidelines, and a discussion of open challenges such as AutoML, LLMs, and federated learning.","tokens_in":34835,"tokens_out":8560,"duration_ms":80135,"significance":"The topic is timely and the paper's architecture is a real strength: it unifies feature selection and generation under a single taxonomy, distinguishes traditional, RL, and generative approaches, and gives explicit formalizations in Section 5. The comparative discussion in Section 6 and the practical guidelines in Section 6.2 are useful and broadly accurate at a high level. If the citation apparatus were reliable, this would be a valuable reference for researchers entering tabular data-centric AI. The manuscript makes no novel algorithmic claims, so there are no derivations or fitted parameters to verify; its evidentiary burden is entirely attributional, which is exactly where the current version fails. The breadth of coverage is genuine, but the broken citations mean the paper cannot currently serve as the trustworthy comprehensive reference it sets out to be.","major_comments":[{"comment":"The citation apparatus for the filter-based methods is internally inconsistent and contains a placeholder. The text cites the chi-square test as [38, 99], ANOVA as [77, 132, 155], Pearson correlation as [193], and mutual information as [5, 193], while Figure 2 attributes the same methods to [45, 46], [47, 48, 49], [50], and [50, 51], respectively. In addition, the sentence \"Methods like mRMR (Maximum Relevance Minimum Redundancy) [? ]\" contains a literal unresolved placeholder. Because a survey's primary function is to let readers locate and verify the methods it describes, these inconsistencies make the central claim of a reliable comprehensive survey unverifiable as submitted. The authors should audit every in-text citation against the final reference list.","section":"§3.1.1 and Figure 2"},{"comment":"The reference list contains at least three duplicate entries: [32] and [33] are the same Fan et al. paper, [54] and [55] are the same Huang et al. paper, and [91] and [92] are the same Liu et al. paper. These duplicates are not harmless: Section 5.1.1 cites [92] for the Monte Carlo early-stopping method while [91] is the identical paper, and Section 3.1.4 cites [55] for the genetic-algorithm-plus-mutual-information hybrid while [54] is the same work. The list must be deduplicated and all downstream numbering recomputed.","section":"References [32]/[33], [54]/[55], [91]/[92]"},{"comment":"The advanced RL feature-selection entries disagree between text and taxonomy figure. The text attributes the multi-agent framework to Liu et al. [89], the GCN-based collaboration to [90], the group-wise method to Fan et al. [34], IRFS to [35], and decision-tree feedback to [33], whereas Figure 2 lists `Group-Wise Method [8]`, `Advanced Statistical Summaries & GCNs State Representation [39]`, `Enhanced Reward Scheme [173]`, `Interactive Reinforced Feature Selection [7]`, `Pre-Filtering And Iterative RL [177]`, and `Decision Tree Based [178]`. Some of these entries point to papers unrelated to the cited method; for example, [8] is a pattern-recognition textbook and [39] is a forest-optimization paper. The same problem affects the LLM branch, where the text cites [189] and [188] but Figure 2 cites [190] and [191]. This makes it impossible to verify the survey's coverage of advanced methods.","section":"§5.1.1 and Figure 2"},{"comment":"Figure 2 promises a `Latent Representation Learning` category under traditional feature transformation, with subcategories for autoencoders, categorical embeddings, and PCA, but Section 4, the corresponding textual treatment, contains no such subsection. PCA appears only later as a dimensionality-reduction suggestion in Section 6.2, and autoencoders appear only in the feature-selection discussion, e.g., SDAE-LSTM in Section 3.1.3. Either a substantive subsection is missing or the taxonomy overstates the coverage; the paper should align the figure with the text before claiming comprehensiveness.","section":"Figure 2 vs Section 4"}],"minor_comments":[{"comment":"The phrase `F or feature selection` contains a stray capitalization artifact and should read `For feature selection`.","section":"Section 1"},{"comment":"The figure labels a whole branch `Feature Transformation`, while the text consistently uses `Feature Generation`; please unify the terminology.","section":"Figure 2 and Section 4"},{"comment":"The reward formula for feature selection is split across a line break in a way that could confuse readers; the expression `R(t) = Perf(...) - lambda * |F(t+1)|` should be set on one line with proper punctuation.","section":"Section 5.1"},{"comment":"Several entries, including [41], [56], [154], [174], [188], and [189], are arXiv preprints; please mark them clearly as preprints and include access dates.","section":"Reference list"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the citation problems are extensive and appear to stem from renumbering after revisions; a simple copyedit will not suffice. I would recommend making acceptance conditional on a complete audit in which every bracketed citation is checked against the final reference list, all duplicate references are removed, and Figure 2 is regenerated from the same source map as the text. The authors should also resolve the Figure 2 `Latent Representation Learning` gap. These issues are within the manuscript's scope and should be fixable, which is why I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe taxonomy here is genuinely useful. Putting single-view vs. multi-view feature selection next to traditional, RL, and generative feature generation in one frame is a service, and the recent RL/generative sections cover work that is otherwise scattered across KDD, ICDM, and arXiv. The comparative tables and the MDP formulation in Section 5 are clear and correct at the level of description. If you need a quick map of who did what in RL/generative feature engineering over the last few years, this survey gets you most of the way there.\n\nBut the citation apparatus is broken in ways that are load-bearing for a survey. The stress-test note checks out on the full text. The chi-square test is [38,99] in Section 3.1.1 but [45,46] in Figure 2; ANOVA, Pearson correlation, and mutual information show the same text-versus-figure conflict. mRMR literally carries a '[?]' placeholder. The reference list itself contains duplicates: [32]/[33], [54]/[55], and [91]/[92] are the same papers. A survey's entire value is that the reader can trace a described method to a verifiable source, and here that chain is broken pervasively. This is not cosmetic; it means the central claim of being a comprehensive, reliable reference is currently unverifiable.\n\nAlso worth noting as minor: Figure 1 lists missing-data imputation as part of data-centric AI, but the survey never covers it, which is a small overreach. And the advanced-methods sections lean heavily on the authors' own prior work; that is not a flaw by itself, but it does make the 'comprehensive' claim sit a bit loosely in the RL/generative chapters.\n\nWho is this for? A researcher or practitioner new to automated feature engineering who wants an organized map with pointers. They will find the map helpful, but they cannot trust the pointers until the citation-reference mapping is fixed, the reference list is deduplicated, and the '[?]' is resolved. The paper deserves a serious referee, but the referee's job right now is to send it back for a rigorous citation cleanup, not to judge the science. The conceptual organization is coherent and the effort is honest; the execution is what failed.\n\nI would not cite it in its current state. After a careful revision, I would.\n\nRecommendation: send to peer review with the expectation of major revision focused on verification of every citation against the reference list.","headline":"A useful and current survey frame for tabular feature selection and generation, but the broken citation-reference machinery undercuts its reliability as a reference work and needs a serious pass before it can serve its purpose.","tokens_in":35391,"tokens_out":1580,"would_cite":false,"duration_ms":18892,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that tabular data-centric AI reduces to two data transformations — feature selection and feature generation — and that one taxonomy can organize traditional, reinforcement-learning, and generative methods for both.","keywords":["data-centric AI","tabular data","feature selection","feature generation","reinforcement learning","generative AI","automated feature engineering","survey"],"falsifier":"A reader can test the survey's utility claim directly by resolving every citation marker in Section 3.1.1 and Figure 2 against the reference list; broken markers — such as the maximum-relevance minimum-redundancy method cited only as '[? ]' and figure entries whose numbers do not correspond to the listed references — would falsify the claim that this is a dependable map of the field.","tokens_in":34416,"feed_emoji":"📊","tokens_out":5069,"duration_ms":49189,"temperature":0.7,"pith_summary":"This survey claims that the practical route to better AI on tabular data runs through the data itself, not just the model. It argues that tabular data transformation has two core tasks — feature selection, which keeps useful columns and drops redundant ones, and feature generation, which creates new columns from existing ones — and that all current methods fit into a single taxonomy. The survey places classical filter, wrapper, and embedded selection alongside multi-view variants, then shows how reinforcement learning and generative AI recast both tasks as automated search problems. It closes with a comparison table and practical guidelines. If the synthesis is right, a practitioner can look up any tabular-data pain point and find the family of solutions designed for it.","feed_headline":"Two tasks rule tabular data AI: feature selection and generation","feed_subtitle":"A new survey ties traditional, reinforcement, and generative methods into one framework so teams can pick the right tool.","key_machinery":"The load-bearing object is the survey's taxonomy, summarized in Figure 2, which partitions tabular data-centric AI into feature selection and feature generation, then under each into traditional methods (filter, wrapper, embedded, hybrid, and multi-view) and advanced methods (reinforcement learning and generative AI). Two formal mechanisms carry the advanced half: the Markov decision process formulation, whose state is the current feature set (plus transformation history for generation), whose actions are selecting or deselecting a feature or applying a mathematical operation, and whose reward balances downstream performance against set size or complexity; and the encoder-decoder-evaluator formulation, in which transformation sequences are encoded into a continuous embedding, decoded back to sequences, and evaluated by predicted performance, enabling a gradient-based search for better transformations.","core_discovery":"On its own terms, the paper's central claim is that feature selection and feature generation are the two essential operations of tabular data-centric AI, and that prior surveys covered only part of this picture. It claims to bridge those gaps by integrating traditional methods, reinforcement learning, and generative AI for both operations under one taxonomy. Feature selection is formalized as choosing a subset $S \\subseteq \\{1,\\dots,d\\}$ of the $d$ original features, and feature generation as producing a transformed dataset through mathematical operations. Advanced methods are then formulated as Markov decision processes for reinforcement learning — the state is the current feature set, the action selects or deselects a feature or applies a transformation, and the reward balances downstream performance against penalties — and as an encoder-decoder-evaluator architecture for generative AI, where transformation sequences are mapped to a continuous embedding space, reconstructed, and scored so that better transformations can be found by gradient-based search. The paper also identifies open challenges, including scalability, interpretability, privacy-preserving feature engineering, and LLM-based generation.","pith_inferences":["If the unified formulations are taken seriously, a natural next step the paper does not take is a shared benchmark that scores feature-selection and feature-generation methods on identical tabular datasets, which would make the taxonomy testable rather than descriptive.","The encoder-decoder-evaluator pattern suggests a testable extension: pretraining the embedding space on synthetic transformation records from many datasets should improve downstream feature quality on a held-out tabular task, and failure would indicate the embedding is not capturing transferable knowledge.","The survey's emphasis on interpretability points to an under-explored hybrid: generative methods that produce candidate features but require a domain expert's validation before entering the model, potentially combining scalability with trust.","The LLM-based directions imply that text-informed feature generation will matter most in domains where tabular rows come with side text, such as medical records or product descriptions; a controlled comparison against text-free generation on such data would quantify the gain."],"forward_implications":["A practitioner facing a high-dimensional table can use the taxonomy to place their problem: cheap statistical screening first, wrapper or embedded selection when feature interactions matter, reinforcement learning for sequential or dynamic feature decisions, and generative models for transferring feature knowledge across tasks.","Feature selection and feature generation, usually treated as separate pipelines, can be viewed as one data-space refinement problem with a shared optimization language.","Reinforcement-learning-generated feature transformation records can be compressed into an embedding space, meaning experience from many datasets can be reused instead of rediscovered.","LLM-based feature generation can operate without a separate learned predictor, using in-context learning and retrieval-augmented generation from external knowledge to create explainable features.","The comparison of traditional and advanced methods gives concrete guidance: prefer traditional methods for small, static, interpretability-critical datasets and advanced methods for high-dimensional, dynamic, or multimodal ones."],"supporting_citations":[{"why":"The prior feature-selection survey whose focus on traditional methods defines the gap the paper extends.","marker":"[80]"},{"why":"A feature-selection-only survey used to show that prior work overlooked feature generation.","marker":"[13]"},{"why":"Another feature-selection review cited as neglecting generation and RL/generative advances.","marker":"[136]"},{"why":"A feature-engineering survey for unstructured data, the contrast that motivates the tabular focus.","marker":"[103]"},{"why":"Multi-agent reinforcement learning for feature selection, the anchor example for RL-based selection in the advanced-methods section.","marker":"[89]"},{"why":"Reinforcement-enhanced autoregressive feature transformation (MOAT), a core generative feature-generation method used in the comparison.","marker":"[148]"},{"why":"GAINS, the deep encoder-decoder-evaluator framework that anchors the generative feature-selection section.","marker":"[165]"},{"why":"DIFER, the early differentiable AutoML feature-generation method that later generative work builds on or critiques.","marker":"[197]"}],"fun_headline_variants":["Feature selection + generation: the two pillars of tabular AI","One survey to rule tabular data prep: select, generate, repeat","From RL to generative: unifying tabular feature engineering","Data-centric AI: mastering feature selection and generation","Tabular data made smarter: the survey that ties it all together"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The survey's usefulness rests on its descriptions and citations of prior work being accurate and correctly attributed, because a reader who cannot trust the mapping from method to citation cannot act on the taxonomy.","fun_headline_variants_meta":{"raw":{"variants":["Feature selection + generation: the two pillars of tabular AI","One survey to rule tabular data prep: select, generate, repeat","From RL to generative: unifying tabular feature engineering","Data-centric AI: mastering feature selection and generation","Tabular data made smarter: the survey that ties it all together"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001584,"raw_usage":{"total_tokens":6296,"prompt_tokens":903,"completion_tokens":5393,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":5309}},"tokens_in":519,"tokens_out":5393,"duration_ms":31766,"temperature":1.0,"reasoning_tokens":5309,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:07:08.420716+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader can test the survey's utility claim directly by resolving every citation marker in Section 3.1.1 and Figure 2 against the reference list; broken markers — such as the maximum-relevance minimum-redundancy method cited only as '[? ]' and figure entries whose numbers do not correspond to the listed references — would falsify the claim that this is a dependable map of the field.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Another feature-selection review cited as neglecting generation and RL/generative advances."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Reinforcement-enhanced autoregressive feature transformation (MOAT), a core generative feature-generation method used in the comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DIFER, the early differentiable AutoML feature-generation method that later generative work builds on or critiques."}],"review_version":1}