{"id":"4254898a-170a-4030-a92f-e233f01e6d8b","arxiv_id":"2501.07096","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"IDCLRec disentangles user behaviors into interests and intents and applies intent-item contrastive learning to improve sequential recommendation accuracy beyond current baselines.","lead":"This paper introduces IDCLRec, a sequential recommendation model that separates user behavior into stable interests and dynamic intents, then uses item-aware contrastive learning to sharpen intent representations. On three Amazon datasets it reports consistent gains over existing intent-based recommender systems, with the largest improvements on the sparsest dataset.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is not statistically established: gains are 3-seed averages with no error bars or significance tests, and the smallest claimed margin (Beauty NDCG@5, about 0.0023 absolute) is within plausible seed noise.","rationale":"The reader's verdict is CONDITIONAL and I largely agree. I focused on statistical reliability rather than the semantic disentanglement assumption because the strongest claim is explicitly about being the new SOTA: even if the intent/interest decomposition is only a useful inductive bias rather than a true semantic split, the model could still be SOTA, whereas unreported seed variance would undercut the headline numbers regardless of the interpretability story. The paper's own Section 5.4.1 concedes that the threshold hyperparameter may be almost redundant ('introducing an additional hyperparameter that is threshold may be unnecessary'); that is a minor internal limitation, not a threat to the empirical result. The Section 5.5 t-SNE separation is circular as evidence for disentanglement, but circularity affects interpretation, not the comparison against baselines. The absence of released code is a reproducibility limitation, not evidence that the results are wrong. A paired multi-seed significance check is the single most direct way to resolve whether the reported 4.75% to 14.62% improvements are real and stable; if they survive, the CONDITIONAL verdict can move toward ACCEPT, and if they do not, the SOTA claim should be downgraded.","tokens_in":22399,"tokens_out":10243,"duration_ms":106518,"concrete_test":"Ask the authors to release per-seed scores (or rerun IDCLRec and ICSRec with at least 10 seeds on the same splits) and compute paired bootstrap or Wilcoxon signed-rank tests for HR@20 and NDCG@20 on each of the three datasets, reporting 95% confidence intervals for the relative improvements. If any IDCLRec-versus-ICSRec difference fails to reach significance, the SOTA claim should be weakened; if all 18 comparisons remain significant, the empirical claim is strongly supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Section 5.2 is that IDCLRec outperforms the best baseline (ICSRec) by 5.47% to 14.62% HR and 4.75% to 12.70% NDCG. For this to be a reliable SOTA claim, the differences must be robust to seed variation and evaluation noise. Section 5.1.4 reports only that experiments use three random seeds and averages; no standard deviations, confidence intervals, or significance tests are reported. Some margins are small in absolute terms: Beauty NDCG@5 is 0.0507 for IDCLRec versus 0.0484 for ICSRec, about 0.0023, and this 4.75% relative gain is the lower bound of the paper's own claimed range. With three seeds and no per-seed values, such a margin may be within ordinary run-to-run variance, so 'significantly outperforms' is under-supported. The t-SNE visualization in Section 5.5 does not help here: it is produced from the model's own representations and would neither confirm nor refute a performance difference. The missing code and per-seed logs further prevent independent verification. If the underlying variance is small the concern evaporates; if not, the headline should be softened to a suggestive improvement pending statistical validation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes IDCLRec, a sequential recommendation model that disentangles user behavior representations into stable interests (extracted by a causal cross-attention mechanism) and dynamic intents (modeled as the residual), aggregates similar intents into a user-level categorical intent via importance-weighted attention, and applies two item-aware contrastive losses. The model is evaluated on three Amazon datasets (Sports, Beauty, Toys) against eleven baselines, reporting improvements in HR and NDCG across all settings, with ablations supporting each component.","tokens_in":22677,"tokens_out":5586,"duration_ms":51941,"significance":"If the performance results are reliable, IDCLRec would be a strong new state-of-the-art on these standard benchmarks, and the intent-interest disentanglement idea is a useful conceptual contribution. The paper provides a thorough experimental protocol: three datasets, eleven baselines, ablation studies, and hyperparameter sensitivity analyses, and the method is clearly described. However, the lack of statistical significance testing and the self-referential nature of the disentanglement analysis limit the strength of the central claims.","major_comments":[{"comment":"The claim that IDCLRec 'significantly outperforms' baselines is not statistically substantiated. The paper reports only three-seed averages without standard deviations, confidence intervals, or significance tests. The smallest margin—Beauty NDCG@5, 0.0507 vs. 0.0484 (4.75% relative)—is small enough that it could plausibly lie within run-to-run variance for transformer-based recommenders. Please report per-seed values, include error bars in Tables 3 and 4, and provide paired significance tests (e.g., bootstrap or t-test) across all datasets and metrics. Without this, the central SOTA claim is under-supported.","section":"Section 5.1.4 and Table 3"},{"comment":"The t-SNE visualization does not independently validate the disentanglement of intents and interests. Because I^u is constructed as the residual H^u - R^u and R^u is produced by a different attention mechanism, the two representations are expected to occupy different regions of the latent space by construction. The figure therefore shows that the model separates these two vectors, but not that they correspond to semantically distinct user characteristics (dynamic motivations vs. stable tastes). Please provide external validation, such as evaluating the temporal stability of interest representations, probing intent representations with known item-category or time-window signals, or comparing against a control in which the residual is replaced by a random vector.","section":"Section 5.5 and Eq. (7)"},{"comment":"The training loss for next-item prediction is written as L_rec = -exp(y_hat_g) / sum_i exp(y_hat_i), which is the negative softmax probability, not the cross-entropy loss. The correct cross-entropy is -log(exp(y_hat_g) / sum_i exp(y_hat_i)). As written, the loss is not a proper classification loss and would not behave as standard cross-entropy during training. Please correct the equation (and verify that the implementation matches the corrected version), or clarify if a different loss is intended.","section":"Eq. (20)"},{"comment":"The ablation results are also reported as single numbers without variance. The conclusions in Section 5.3, such as the relative contributions of L_d, L_CL1, and L_CL2, would be more convincing with error bars or significance tests, given that the reported differences between some ablations (e.g., IDCLRec vs. A on Sports NDCG@20: 0.0426 vs. 0.0425) are very small. Please include per-seed standard deviations for the ablation experiments, or at least assess whether the observed differences are stable across seeds.","section":"Tables 3 and 4"}],"minor_comments":[{"comment":"The paper itself notes that the threshold delta may be unnecessary and suggests setting it to 0 to consider all intents. This observation undermines the design choice of introducing an extra hyperparameter; please either remove the threshold from the final model or provide a stronger justification for keeping it.","section":"Section 5.4.1"},{"comment":"The construction of the augmented sequence S_u' is unclear: the paper describes it as 'the interaction history of user u' with the target item as v_{N+1}', but does not specify how such a sequence is obtained from the sequence-splitting procedure. Please define the augmentation operation explicitly.","section":"Section 4.6.1"},{"comment":"No code or per-seed training logs are provided, which prevents independent verification of the reported results and the claimed standard deviations (if any). Please release code and seeds to facilitate reproducibility.","section":"General"},{"comment":"The text refers to Figure 2-(a), (b), (c) but the figure caption does not label these subpanels; please add subfigure labels for clarity.","section":"Section 3, Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical contribution but the central SOTA claim requires statistical support. The t-SNE-based disentanglement analysis is not convincing as evidence, and the training-loss equation appears to be a typo that must be fixed. The paper's own admission that the threshold hyperparameter may be unnecessary suggests the model could be simplified. The work is within the scope of the journal, but the revisions above are essential before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Basically: IDCLRec is a serious empirical paper with a plausible new combination of ideas, and it probably does beat the baselines. But the paper's strong SOTA claim rests on three-seed averages with no error bars, and its central disentanglement evidence is t-SNE on the model's own representations.\n\nWhat's new is the combination of intent-interest disentanglement with a user-specific categorical intent module and two item-aware contrastive losses. The causal cross-attention interest extractor and the residual intents are a clean way to separate stable from dynamic factors. The importance-weighted attention over similar intents is a sensible alternative to fixed-K clustering, and the ablation suite is thorough: seven variants, showing each component contributes, with the largest drop when disentanglement is removed entirely. That suggests the decomposition is doing real work, not just adding parameters.\n\nThe soft spots are real but not fatal. The statistical support is the main issue: three seeds, averages only, no standard deviations or significance tests. The smallest absolute margin (Beauty NDCG@5: 0.0507 vs 0.0484, about 0.0023) is well within plausible seed noise. Some margins are larger and more convincing on their face, but without per-seed values the reader cannot judge. The t-SNE in Section 5.5 does not help because it is generated from the model's own representations; it cannot independently confirm that intents and interests are semantically distinct. Also, the 'residual equals intent' assumption (Eq. 7) could be absorbing noise rather than genuine intent, and the performance gain could come from regularization instead of true disentanglement. That is a legitimate alternative explanation they do not test. Missing code makes independent verification harder.\n\nThis paper is for people working on intent-based or contrastive sequential recommendation, and for practitioners wanting a few percent gain on standard Amazon benchmarks. It deserves a serious referee: the experiments are broad, the ablations are informative, and the core result is a consistent improvement across three datasets and six metrics. I would ask for per-seed numbers, significance tests or at least variance reporting, a non-t-SNE check of disentanglement, and code release before acceptance.\n\nRecommendation: send to peer review, with the statistical validation and disentanglement evidence as required revisions.","headline":"A well-built empirical model that likely beats SOTA on sequential recommendation, but the headline margin is not statistically pinned down and the disentanglement evidence is self-referential.","tokens_in":23185,"tokens_out":2075,"would_cite":false,"duration_ms":20981,"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":"IDCLRec claims that separating stable user tastes from dynamic intents improves sequential recommendation, and that the residual after causal cross-attention is the right place to model intents.","keywords":["sequential recommendation","intent modeling","interest disentanglement","causal cross-attention","item-aware contrastive learning","importance-weighted attention","next-item prediction"],"falsifier":"Generate synthetic sequences from a known generative process in which each user has a fixed interest vector and a piecewise-constant intent vector that switches at known timestamps; train IDCLRec and measure whether the residual intent representations are more similar within an intent block than across blocks and whether the inferred switch points align with the injected ones. If the residual does not track the injected intents, the claim that $I^u = H^u - R^u$ encodes dynamic intent is refuted.","tokens_in":22184,"feed_emoji":"🎯","tokens_out":9413,"duration_ms":86739,"temperature":0.7,"pith_summary":"This paper tries to establish that sequential recommendation models should stop treating user interaction histories as a single behavioral signal. It argues that those histories mix stable personal tastes with evolving motivations, and that only by separating them can a model understand and predict user choices. The proposed model, IDCLRec, performs the separation by using causal cross-attention to extract interests and treating everything left over as intents, then improves intents with item-aware contrastive learning. The paper reports that this design outperforms all compared baselines on three real-world e-commerce datasets, with gains in hit rate and ranking quality that are largest on the sparsest dataset. If correct, the result would make two-factor behavior decomposition a standard component of intent-based sequential recommenders.","feed_headline":"Splitting taste from intent lifts recommendation accuracy","feed_subtitle":"A disentangled intent-interest model beats the prior best by 14.6% in hit-rate and 12.7% in ranking quality.","key_machinery":"The load-bearing object is the residual identity $I^u = H^u - R^u$, which defines intents as whatever remains after the causal cross-attention module extracts stable interests from the sequence representations. The causal cross-attention uses a causal mask so each current behavior attends only to previous behaviors, and layer normalization after the residual connection retains only features shared between current and previous steps, forcing $R^u$ to encode consistent patterns. The similarity-adjustment loss $\\mathcal{L}_d = \\frac{1}{N-1}\\sum_{t=1}^{N-1}(1 - \\mathrm{sigmoid}(\\mathrm{sim}(i^u_t, i^u_{t+1})))$ enforces gradual intent drift, and the importance-weighted attention module selects intents whose similarity to the most recent intent exceeds a threshold $\\delta$, computes sigmoid relevance scores between them, and softmax-normalizes summed relevance to weight each intent's contribution to the final categorical intent vector. Two InfoNCE contrastive objectives connect intents to items: $\\mathcal{L}_{CL1}$ aligns intents of sequences sharing the same target item, and $\\mathcal{L}_{CL2}$ aligns each intent with the averaged embedding of items interacted with under that intent. These mechanisms together carry the argument from raw sequences to the final user representation $h^u = r_N^u + i^u$.","core_discovery":"IDCLRec's central claim is that user behavior representations are composed of two separable signals: stable interests and dynamic intents. The model extracts interests with a causal cross-attention mechanism that aggregates only previous behaviors relevant to the current one, then adds the current behavior back with layer normalization to form $R^u = \\mathrm{LayerNorm}(\\hat{R}^u + H^u)$. Intents are defined as the residual $I^u = H^u - R^u$. A similarity-adjustment loss makes adjacent intents evolve smoothly, and an importance-weighted attention module aggregates previous intents most similar to the most recent intent into a user-specific categorical intent without any preset number of categories. Two item-aware contrastive losses further sharpen this intent representation: one aligns intents from sequences whose next target item matches, and the other pulls an intent toward the average embedding of the items interacted with under that same intent. On the Sports, Beauty, and Toys datasets, the paper reports the highest hit rate and ranking quality across all baselines, with 5.47% to 14.62% improvements in HR and 4.75% to 12.70% improvements in NDCG over the best baseline.","pith_inferences":["Inference: the causal cross-attention mechanism can be viewed as a smoothing operation over behavior vectors, so part of the improvement attributed to disentanglement may be regularization that stabilizes representations rather than a faithful split into taste and intent.","Inference: a decisive external test of the split would use synthetic sequences generated from known piecewise-constant intents and fixed user interest vectors, checking whether the residual intent representations switch at the injected points; such validation is absent from the paper.","Inference: the same interest/intent decomposition could be transplanted to session-based recommendation, where there is no persistent user identity, by treating session-level recurring themes as interests and transition-level purposes as intents.","Inference: the threshold $\\delta$ may be unnecessary; the paper's own $\\delta=0$ test shows nearly identical performance, so dropping the threshold and keeping only importance-weighted attention would produce a simpler model with one fewer hyperparameter."],"forward_implications":["If the paper is right, models that claim to capture user intents should be compared against a two-factor baseline that separates stable interests from dynamic intents, since the ablations attribute a large part of the gain to this separation.","Fixed global numbers of intent categories can be replaced by per-user similarity-based aggregation; the paper shows performance remains stable across thresholds, suggesting the threshold itself may be removable.","Explicitly tying intents to the items that express them through contrastive learning becomes a default ingredient: both same-target-item alignment and intent-to-item-combination alignment improve ranking in the reported experiments.","The larger relative gains on the sparser Sports dataset suggest that the disentanglement and item-aware contrastive objectives help most when user histories are short, a setting where sequential recommenders usually struggle."],"supporting_citations":[{"why":"It supplies the SASRec transformer encoder that produces the sequence behavior representations which the model then separates.","marker":"[20]"},{"why":"It establishes the intent contrastive learning line and the fixed-K intent category assumption that IDCLRec is designed to replace, and it is a comparison baseline.","marker":"[7]"},{"why":"It is the state-of-the-art intent baseline whose learned intents are analyzed to motivate user-level importance weighting and item-aware contrastive learning, and its target-item augmentation is extended.","marker":"[35]"},{"why":"It provides the importance extraction idea of sigmoid relevance scores summed and softmax-normalized, which IDCLRec adapts to weight similar intents.","marker":"[33]"},{"why":"It defines the InfoNCE contrastive objective used for both the intent-intent and intent-item contrastive losses.","marker":"[32]"},{"why":"It introduces same-target-item supervised positive sampling and sequence splitting augmentation, the augmentation used for the intent-intent contrastive task.","marker":"[36]"},{"why":"It supplies the Amazon e-commerce review dataset on which the reported experiments and baseline comparisons run.","marker":"[15]"},{"why":"It supplies the Amazon review data underlying the Sports, Beauty, and Toys subsets.","marker":"[31]"}],"fun_headline_variants":["Disentangling taste from intent boosts recommendations","Intent-interest split sharpens sequential recommendations","Separate stable taste from fleeting intent for better recs","Item-aware contrastive learning improves intent modeling","Breaking up intent and interest lifts hit rate 14.6%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that subtracting the extracted stable-interest component from behavior vectors leaves genuine dynamic intents; if that residual is mostly noise or still contains interest information, the performance gains could come from extra training regularization rather than from true intent disentanglement.","fun_headline_variants_meta":{"raw":{"variants":["Disentangling taste from intent boosts recommendations","Intent-interest split sharpens sequential recommendations","Separate stable taste from fleeting intent for better recs","Item-aware contrastive learning improves intent modeling","Breaking up intent and interest lifts hit rate 14.6%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000162,"raw_usage":{"total_tokens":1304,"prompt_tokens":1078,"completion_tokens":226,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":694,"completion_tokens_details":{"reasoning_tokens":163}},"tokens_in":694,"tokens_out":226,"duration_ms":2580,"temperature":1.0,"reasoning_tokens":163,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:48:42.362684+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate synthetic sequences from a known generative process in which each user has a fixed interest vector and a piecewise-constant intent vector that switches at known timestamps; train IDCLRec and measure whether the residual intent representations are more similar within an intent block than across blocks and whether the inferred switch points align with the injected ones. If the residual does not track the injected intents, the claim that $I^u = H^u - R^u$ encodes dynamic intent is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It establishes the intent contrastive learning line and the fixed-K intent category assumption that IDCLRec is designed to replace, and it is a comparison baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It provides the importance extraction idea of sigmoid relevance scores summed and softmax-normalized, which IDCLRec adapts to weight similar intents."}],"review_version":1}