{"id":"1752a15b-a0a3-4186-9ff0-ecd6dac73a55","arxiv_id":"2411.18631","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ClardRec uses counterfactual-style score comparisons to split item representations into query-related and general parts, and transfers the general parts to improve recommendation.","lead":"A new recommender-system framework, ClardRec, tries to separate the parts of a product or video that match a user's search query from the parts that reflect general taste, so only the general parts are transferred to recommendation. The method reports small but consistent gains over existing search-enhanced recommenders on one public and one industrial dataset.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 13's triplet loss appears to maximize the opposite of the inequalities in Eqs. 10-12, making the central counterfactual disentanglement mechanism internally inconsistent as written.","rationale":"The reader's weakest_assumption identifies the unverified inequalities in Eqs. 10-12. My concern is more specific: even granting those inequalities, the objective constructed to enforce them has the wrong sign. This is the single most load-bearing issue because every other component of the paper depends on i_src/q being a query-independent representation learned from counterfactual supervision. If L_cd rewards the opposite ordering, the feature augmentation and data augmentation modules are transferring a representation whose semantics are not established by the stated loss. The ablations and qualitative visualizations do not resolve this: -CD shows only that removing the extra terms hurts accuracy, which is compatible with the terms acting as generic regularization; the t-SNE plot and case study are consistent with many alternative mechanisms. I therefore cannot sign on to a conditional acceptance contingent on code and data, because releasing code would only reveal which of the contradictory formulations was actually implemented. As written, the central mechanism is internally inconsistent, so the appropriate outcome is rejection of the current manuscript; a corrected version with a sign-consistent objective and released code could be reconsidered.","tokens_in":19483,"tokens_out":9501,"duration_ms":96768,"concrete_test":"Algebraic check: instantiate scalar scores with Delta_q_i/q = -3 and Delta_u_i/q = -1, which satisfy Eq. 10 under the paper's sign convention, and compare -log sigma(Delta_q_i/q - Delta_u_i/q) with -log sigma(Delta_u_i/q - Delta_q_i/q). The former is larger, confirming that Eq. 13 pushes toward Delta_q_i/q > Delta_u_i/q. If confirmed, obtain the released code and inspect the implemented loss, or re-run the KuaiSAR experiments with the arguments of Eq. 13/14 swapped to match Eqs. 10-12; if the corrected variant does not reproduce the reported results or performs worse, the published equations do not describe the evaluated system and the disentanglement claim is unsupported.","verdict_should_be":"REJECT","load_bearing_attack":"The load-bearing supervision for disentanglement is the claim that removing query-related features changes query-item scores more than user-item scores (Eq. 10), with symmetric statements in Eqs. 11-12. Under the paper's own sign convention, Delta = score_after - score_before, and the text states that for positive items Delta <= 0 and for negative items Delta >= 0 (Section 4.2.2). For a positive item, Eq. 10 with both Deltas negative is equivalent to Delta_q_i/q < Delta_u_i/q, i.e., Delta_q_i/q - Delta_u_i/q < 0. The first term of L/q_cd in Eq. 13, -log sigma(Delta_q_i/q - Delta_u_i/q), is minimized by making that argument as large as possible, hence by Delta_q_i/q > Delta_u_i/q: the opposite inequality. The negative-item term of Eq. 13 is likewise inverted: for negative items Eq. 10 requires Delta_q_i-/q > Delta_u_i-/q, while the loss pushes Delta_u_i-/q - Delta_q_i-/q > 0. Equations 14 are inverted in the same way for the removal-wise signals. Therefore L_cd as written does not implement Eqs. 10-12; it appears to reward the exact opposite ordering. Consequently, the query-independent representation i_src/q is not supervised by the stated counterfactual signals, and the reported gains cannot be attributed to the claimed disentanglement mechanism. This is an internal inconsistency, independent of whether the inequalities are semantically correct.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ClardRec, a search-enhanced recommendation framework that uses search queries to disentangle item representations into query-related and query-independent components. The query-independent component is then transferred to the recommendation domain through feature augmentation (a gated fusion network) and data augmentation (confidence-weighted contrastive/auxiliary losses). The method is evaluated on KuaiSAR and an industrial E-commerce R&S dataset under both collaborative filtering and sequential recommendation backbones, with consistent reported improvements over baselines.","tokens_in":19857,"tokens_out":8294,"duration_ms":74817,"significance":"If the disentanglement mechanism worked as claimed, the paper would be a useful practical contribution to search-enhanced recommendation, addressing a real negative-transfer problem. The empirical protocol has strengths: recommendation metrics are held out, 99 negative items are used, five random seeds are reported, and multiple backbones and recent baselines are considered. However, the central supervision signal has a sign inconsistency that inverts the stated score-variation inequalities, and the disentanglement is not validated against any external ground truth. Until those issues are resolved, the reported gains cannot be attributed to the claimed counterfactual disentanglement mechanism.","major_comments":[{"comment":"With Delta defined as after-before in Section 4.2.2, the triplet losses in Eqs. (13) and (14) reward the opposite of the inequalities in Eqs. (10)-(12). For a positive item, Eq. (10) states |Delta_q i/q| > |Delta_u i/q|; combined with the paper's own sign statement Delta_q i/q <= 0 and Delta_u i/q <= 0, this is equivalent to Delta_q i/q - Delta_u i/q < 0. The first term of L/q_cd, -log sigma(Delta_q i/q - Delta_u i/q), is minimized by making Delta_q i/q - Delta_u i/q as large as possible, i.e., Delta_q i/q > Delta_u i/q. The negative-item term is inverted in the same way: Eq. (10) for a negative item implies Delta_q i-/q > Delta_u i-/q, while the loss maximizes Delta_u i-/q - Delta_q i-/q. The same reversal affects L/u_cd in Eq. (13) and both terms of Eq. (14) against Eq. (12). Consequently, L_cd as written does not implement the stated counterfactual supervision, and the query-independent representation i_src/q is not trained to satisfy the assumed ordering. The ablation results in Table 2 therefore cannot be attributed to the claimed disentanglement mechanism.","section":"Section 4.2.2, Eqs. (10)-(14)"},{"comment":"The disentanglement is supervised only by the assumed score-variation orderings; there is no external check that i_src/q actually encodes query-independent, recommendation-relevant item semantics. The t-SNE visualization in Figure 3 and the case study in Figure 4 are qualitative illustrations, not quantitative measurements. I recommend adding a probe test, such as a classifier predicting query words or item categories from i_src/q versus i_src/u, or an ablation in which the gated split is replaced by a random split with a random score-ordering objective, to demonstrate that the gains come from genuine disentanglement rather than from added model capacity or generic regularization.","section":"Section 4.2.2, Eqs. (10)-(12) and Section 5.4"}],"minor_comments":[{"comment":"The notation i_emb/q and i_emb/u is confusing: i_emb/q is defined as the embedding after removing query-related features, yet the '/q' subscript more naturally suggests query-related content. Please rename the two components (e.g., i_gen and i_qry) and state the convention explicitly at first use.","section":"Section 4.2.1"},{"comment":"The asterisks are declared to indicate statistical significance over the best baseline, but several bold entries (e.g., MLP+ClardRec on KuaiSAR HR@1) are not starred; please clarify when differences are significant and specify the number of seeds and the exact test procedure.","section":"Section 5.1.3 and Table 1"},{"comment":"The t-SNE visualization is presented as evidence of successful disentanglement, but no quantitative separation or alignment measure is reported; adding a distance- or centroid-based metric would make the answer to Q3 more convincing.","section":"Section 5.4 and Figure 3"},{"comment":"Reference [31] appears garbled ('K. Z. Zhang P.K. Kannan Y Yang'); please correct the author list.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The sign inconsistency in Eqs. (13)-(14) is the deciding issue. I chose major_revision rather than reject because the error is local and correctable, but the authors must rerun the experiments with the corrected objective and provide quantitative evidence for the disentanglement claim. Without those, the paper should not be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the central counterfactual disentanglement loss in ClardRec looks inverted. The stress-test note holds up. For positive items, Eq. 10 demands |Δ_qi/q| > |Δ_ui/q|; with both deltas negative under the paper's sign convention, this means Δ_qi/q < Δ_ui/q, so Δ_qi/q - Δ_ui/q < 0. But the first term of Eq. 13 is -log σ(Δ_qi/q - Δ_ui/q), minimized by making that argument positive. The negative-item term and the removal-wise losses in Eq. 14 are inverted the same way. So L_cd rewards the opposite ordering from the inequalities it claims to supervise, and the query-independent representation is not being trained with the stated counterfactual signal. This is not a small gap; it is the paper's load-bearing mechanism.\n\nWhat is genuinely good: the problem is real, and the paper documents it well with the t-SNE and category-rank analyses. The overall framework – gated disentanglement, feature augmentation, confidence-weighted data augmentation – is plausible, and the ablations are consistent and complete enough (removing each module hurts). Four backbones, two datasets including an industrial one, and careful hyperparameter reporting show solid empirical practice. The case study is illustrative rather than proof, but fine.\n\nThe other soft spots are secondary. There is no independent check that the query-independent part actually excludes query-specific attributes; the t-SNE is suggestive but does not verify the semantic split. The industrial dataset has no released code or data, so the strongest results are not reproducible. And the 'counterfactual' framing is really just an assumed inequality about score variations, not derived from a causal model – that would be acceptable if the loss actually enforced it, but it doesn't.\n\nWho is this for? Researchers working on search-enhanced recommendation and disentangled transfer. The empirical setup and the augmentation ideas are worth reading, and the sign error is an instructive cautionary tale. As it stands, I would not cite the method as working as described, but I would want to see a corrected version.\n\nFor peer review: yes, an editor should send it out – the paper is well-structured and the claims are checkable, which is exactly what a referee should catch. But my recommendation would be major revision at best, with the loss fixed, code released, and an independent verification of the disentanglement.","headline":"The paper's core counterfactual triplet loss appears to invert its own inequalities, so the claimed disentanglement mechanism likely does not do what is described.","tokens_in":20345,"tokens_out":4481,"would_cite":false,"duration_ms":40695,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Counterfactual score shifts supervise a split of search-clicked item features into query-specific and general-interest parts, improving recommendation.","keywords":["search-enhanced recommendation","counterfactual learning","representation disentanglement","negative transfer","recommender systems","feature augmentation","data augmentation","sequential recommendation"],"falsifier":"On a dataset where item attributes are labeled as query-specific or general-interest, train ClardRec and inspect whether the query-independent representation $i^{\\mathrm{src}}_{/q}$ preferentially receives the labeled general-interest attributes; the central claim is falsified if those attributes instead land in $i^{\\mathrm{src}}_{/u}$, or if replacing the counterfactual ordering losses with shuffled score-orderings leaves recommendation accuracy unchanged.","tokens_in":19285,"feed_emoji":"🧩","tokens_out":17115,"duration_ms":158229,"temperature":0.7,"pith_summary":"Search-enhanced recommenders use a user's search clicks to learn preferences, but those clicks mix features that match the query being searched with features that reflect the user's general taste. The paper argues that transferring both kinds of features into the recommendation domain leaks search-specific intent and can hurt recommendation accuracy. ClardRec constructs counterfactual supervision from score changes: if removing query-related features changes query-item scores more than user-item scores, and removing general-interest features changes user-item scores more, those relative changes can teach an attribute-aware gating network to split item representations. The query-independent parts are then transferred via feature and data augmentation. Experiments on a short-video dataset and an industrial e-commerce dataset report accuracy gains across collaborative filtering and sequential recommendation backbones.","feed_headline":"Counterfactual signals split search clicks into taste and intent","feed_subtitle":"Score changes teach a model which item features are search-specific and which reflect general interest, improving recommendations.","key_machinery":"The load-bearing mechanism is the triplet counterfactual objective built on score-variation ordering. The model computes user-item preference and query-item match scores, $\\hat{s}_{ui}=u^{\\mathrm{src}}\\cdot i^{\\mathrm{src}}$ and $\\hat{s}_{qi}=q^{\\mathrm{src}}\\cdot i^{\\mathrm{src}}$, and the corresponding scores $\\hat{s}_{ui/q},\\hat{s}_{qi/q},\\hat{s}_{ui/u},\\hat{s}_{qi/u}$ after replacing $i^{\\mathrm{src}}$ with the disentangled representations $i^{\\mathrm{src}}_{/q}$ and $i^{\\mathrm{src}}_{/u}$. The supervision is the assumed ordering $|\\Delta_{qi/q}|>|\\Delta_{ui/q}|$ and $|\\Delta_{ui/u}|>|\\Delta_{qi/u}|$, together with $|\\Delta_{qi/q}|>|\\Delta_{qi/u}|$ and $|\\Delta_{ui/u}|>|\\Delta_{ui/q}|$, where each $\\Delta$ is the corresponding score difference (e.g., $\\Delta_{qi/q}=\\hat{s}_{qi/q}-\\hat{s}_{qi}$). These inequalities become log-sigmoid triplet losses and constraint losses that train an attribute-aware gated network to split each search item embedding into a query-independent part and a query-related part. The query-independent part is then transferred through a gated fusion network for feature augmentation and a confidence-weighted cross-entropy loss for data augmentation.","core_discovery":"The paper's central claim is that the ordering of score changes under counterfactual feature removal is a usable supervision signal for disentangling item representations without feature-level labels. For a clicked (user, query, item) triple, removing the query-related part of the item (leaving the query-independent representation $i^{\\mathrm{src}}_{/q}$) should change the query-item match score more than the user-item preference score, while removing the general-interest part (leaving the query-related representation $i^{\\mathrm{src}}_{/u}$) should change the user-item preference score more. These matching-wise inequalities are paired with removal-wise inequalities comparing the two removal operations, and all are encoded as log-sigmoid triplet losses and score-constraint losses. The same attribute-aware gated network that produces the split also yields gate weights per item attribute, so attributes like category can carry different tendencies for query matching versus general interest. The paper claims that the query-independent representation obtained this way reflects user general interest and that transferring it through feature and data augmentation improves recommendation accuracy.","pith_inferences":["Editorial inference: the score-variation recipe does not depend on search specifically; the same inequalities could treat any contextual variable (session, campaign, device) as the 'query', yielding a generic way to separate context-bound intent from stable preference.","Editorial inference: a direct validity check the paper does not report is attribute-level inspection—comparing gate weights for known search-specific versus general attributes—which would show whether the counterfactual inequalities identify semantically meaningful features or just score sensitivity.","Editorial inference: transfer could go both directions; the disentangled representations learned for recommendation items could in turn be used to explain search results, surfacing the general-interest versus query-intent weights behind each match."],"forward_implications":["Platforms with overlapping search and recommendation identities can obtain general-interest item representations without hand-labeled feature annotations.","The same feature augmentation can be grafted onto different collaborative filtering and sequential backbones, so the benefit is not tied to one architecture.","Confidence-weighted data augmentation down-weights search clicks that mostly reflect query intent, which is the mechanism that mitigates negative transfer from search data.","Because the method requires overlapping users between the two domains, its gains in the reported settings are conditional on that overlap; the paper names this as its main limitation."],"supporting_citations":[{"why":"It supplies the contrastive-learning disentanglement baseline that ClardRec is compared against and whose shared-interest framing the paper refines.","marker":"[23]"},{"why":"It supplies the joint search-recommendation baseline and the shared-embedding design the paper adopts.","marker":"[33]"},{"why":"It supplies the causal search-data baseline that directly transfers search behavior, the approach ClardRec argues suffers negative transfer.","marker":"[22]"},{"why":"It supplies the KuaiSAR dataset, one of the two real-world benchmarks used in the experiments.","marker":"[25]"},{"why":"It grounds the information-retrieval belief that a click under a query depends on query relevance plus user-specific general appeal.","marker":"[3]"},{"why":"It provides a second grounding reference for the same click-generation belief from personalized search on the web.","marker":"[16]"},{"why":"It grounds the product-search setting in which query match and user interest are separable signals, the premise behind the disentanglement split.","marker":"[19]"}],"fun_headline_variants":["Counterfactual scores untangle search intent from user preference","Separate search-specific chaff from general-interest wheat in clicks","Clean search signals: counterfactual tricks to separate taste from intent","Search clicks lie: counterfactual shifts reveal true user taste"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the relative sizes of score changes—query-item scores moving more when query-related features are removed, user-item scores moving more when general-interest features are removed—correctly identify which item features are search-specific and which are general taste, since the learned split is never checked against ground-truth feature labels.","fun_headline_variants_meta":{"raw":{"variants":["Counterfactual scores untangle search intent from user preference","Separate search-specific chaff from general-interest wheat in clicks","Clean search signals: counterfactual tricks to separate taste from intent","Search clicks lie: counterfactual shifts reveal true user taste"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002435,"raw_usage":{"total_tokens":9387,"prompt_tokens":1007,"completion_tokens":8380,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":623,"completion_tokens_details":{"reasoning_tokens":8309}},"tokens_in":623,"tokens_out":8380,"duration_ms":57014,"temperature":1.0,"reasoning_tokens":8309,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:46:29.005233+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a dataset where item attributes are labeled as query-specific or general-interest, train ClardRec and inspect whether the query-independent representation $i^{\\mathrm{src}}_{/q}$ preferentially receives the labeled general-interest attributes; the central claim is falsified if those attributes instead land in $i^{\\mathrm{src}}_{/u}$, or if replacing the counterfactual ordering losses with shuffled score-orderings leaves recommendation accuracy unchanged.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the contrastive-learning disentanglement baseline that ClardRec is compared against and whose shared-interest framing the paper refines."},{"cited_title":"Bruce Croft","cited_arxiv_id":null,"evidence_quote":"It supplies the joint search-recommendation baseline and the shared-embedding design the paper adopts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the causal search-data baseline that directly transfers search behavior, the approach ClardRec argues suffers negative transfer."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It grounds the information-retrieval belief that a click under a query depends on query relevance plus user-specific general appeal."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It provides a second grounding reference for the same click-generation belief from personalized search on the web."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It grounds the product-search setting in which query match and user interest are separable signals, the premise behind the disentanglement split."}],"review_version":1}