{"id":"2b525746-2815-4efc-ad63-017e7b1b0e59","arxiv_id":"2502.02232","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"COPF combines stage-wise constrained graph convolution with a decoupled multi-expert prediction module and reports state-of-the-art results on Beibei, Taobao, and Tmall.","lead":"The paper proposes COPF, a multi-behavior recommendation framework that fuses user behaviors through staged graph constraints and coordinates multi-task prediction with contrastive learning, behavior-fitting experts, and stopped gradients. It reports large gains over prior baselines on three public datasets, making it potentially useful for e-commerce and video recommendation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim depends on an unstated evaluation protocol: no split is specified, and if test target edges enter the B_k adjacency matrices used in Eqs. (5)–(6), the reported gains would be label leakage rather than method quality.","rationale":"The reader's weakest assumption (fixed behavior order) is plausible but less decisive for the target-behavior claim: in Eqs. (5) and (6), the target K is always downstream and its convolution includes all behavior matrices, so buy predictions can still use view/cart/buy edges even if the prescribed order is wrong for some users. The order assumption affects auxiliary representations and the paper's 'pattern' narrative, but it does not by itself invalidate the empirical comparison. The protocol gap is more directly load-bearing: every number in Table 2 depends on it, and the paper gives no way to tell whether test edges enter the graph. That is a correctness risk, not a novelty dispute, and it is checkable from the released code. I therefore keep the CONDITIONAL verdict: acceptance should require a masked-graph reproduction, a stated split/negative-sampling protocol, and per-seed error bars. If the masked reproduction collapses, the proper verdict would be rejection of the SOTA claim.","tokens_in":20963,"tokens_out":12016,"duration_ms":122090,"concrete_test":"Using the released repository, mask the evaluation split: for each user, remove the held-out target (buy) interactions from every B_k before training and before all graph-convolution passes (Eqs. (5), (6), and (11)), then re-run Table 2 on Taobao with the reported hyperparameters. Compare against a transductive rerun in which test positives remain in B_k during propagation but are excluded only from the BPR loss. If masked-graph HR@10 falls below the best baseline (PKEF, 0.1385) or below the claimed 0.1552, the SOTA gain is not attributable to COPF's architecture. Also report per-seed standard deviation over the 5 runs; if the Taobao 12.06% HR improvement is within one standard deviation, the headline claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is empirical: COPF beats 15 baselines by 12–50% HR@10 on three benchmarks. This claim is only as strong as the evaluation protocol, and the protocol is not specified. The manuscript never states how target-behavior test positives are chosen, how negative samples are drawn, whether a validation split is used for the per-dataset hyperparameter searches (loss coefficients λ_k, GCN depth, α), or whether test interactions are removed from the behavior matrices B_k before training and before computing the graph convolutions in Eqs. (5), (6), and (11). The omission is load-bearing because COGCN and the DFME behavior-fitting expert propagate over A_k; if any test positive remains in B_k, the model receives the answer during representation computation. Appendix A.2 only reports behavior-pattern proportions and explicitly says interaction timings are unknown, but it does not describe the split. Thus the reported SOTA cannot currently be separated from a possible label-leakage artifact.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes COPF, a multi-behavior recommendation framework with two components: COGCN, which applies staged behavioral constraints (pre-behavior, in-behavior, post-behavior) to graph-convolutional message passing, and DFME, which combines contrastive alignment, behavior-fitting experts, and stop-gradient operators to mitigate negative transfer in multi-task prediction. On Beibei, Taobao, and Tmall, COPF reports HR@10 improvements of 49.91%, 12.06%, and 24.12% over the best baselines, with ablations and compatibility experiments supporting the contribution of each module. The authors release code at https://github.com/1918190/COPF.","tokens_in":21170,"tokens_out":3304,"duration_ms":34612,"significance":"If the empirical results are reproducible under a clean evaluation protocol, the reported gains are substantial and the ablation and compatibility analyses are unusually thorough for this area. The paper ships code, which is a concrete strength, and the modular design (COGCN and DFME can be studied separately) makes the claims testable. The main weakness is that the central claim is empirical while the evaluation protocol is under-specified; the combinatorial-optimization framing is also not formalized as an optimization problem. The paper does not provide machine-checked proofs or parameter-free derivations, so its value rests almost entirely on the reliability of the experiments.","major_comments":[{"comment":"The evaluation protocol is not specified: the paper does not state how the data are split into train, validation, and test sets, how negative items are sampled for BPR, or whether any hyperparameter selection uses the test set. Most importantly, it never states that test target interactions are removed from the behavior matrices B_k before training and before computing the graph convolutions in Eqs. (5), (6), and (11). Since COGCN and the behavior-fitting experts propagate over A_k, a test positive left in B_k would directly leak the answer into the representation computation. This is load-bearing for the SOTA claim in Table 2. Please specify the exact split (e.g., leave-one-out per user, temporal split, or random split), the negative sampling rule, the validation procedure, and confirm in the text that all test edges are excluded from all adjacency matrices used in training and in the forward pass at evaluation time.","section":"§5.1, Eqs. (5)–(6), (11)"},{"comment":"Per-dataset hyperparameter searches are described only as ranges (e.g., GCN layers in {1,2,3,4}, behavior loss coefficients in {0,1/6,...,1}, scaling factors in {0.1,0.2,0.4,0.8}), but the final per-dataset values are not reported. The paper also reports averages over five runs without standard deviations and marks significance with a star without describing the statistical test. The magnitude of the claimed improvement, especially the 49.91% HR@10 gain on Beibei, cannot be assessed without knowing the selected configuration and the variance across runs. Please report the chosen hyperparameters for each dataset and the standard deviations, and describe the significance test used for the stars.","section":"§5.1.1, Table 2"},{"comment":"The 'combinatorial optimization perspective' is not formalized. The paper counts the number of possible user behavior patterns and states that constraints restrict the solution space, but it never defines the objective function, the decision variables, the feasible set, or the optimization problem being solved. The pre-, in-, and post-behavior constraints are architectural heuristics, and the claimed connection to combinatorial optimization is therefore not established by the text. Either provide a formal statement of the optimization problem with explicit constraints, or rephrase the contribution as a design rationale rather than a solution to a formal combinatorial problem.","section":"§4.2, Definitions 1–5"},{"comment":"The method assumes a fixed, globally shared behavior order (view before cart before buy, or view before collect before cart before buy) while Appendix A.2 explicitly states that the exact timing of user-item interactions is unknown, so the assumed behavior patterns are not directly observed. If the assumed order is violated for a substantial fraction of users, the constraints in Definition 4 and Eq. (5) will discard useful signal. The paper should validate this assumption empirically, for example by comparing with a reversed behavior order, a variant that allows multiple orders, or by reporting the proportion of user histories that are consistent with the assumed order. Without such evidence, part of the reported gain could be an artifact of the ordering assumption rather than a general property of the method.","section":"Appendix A.2, Definition 2 and Eq. (5)"}],"minor_comments":[{"comment":"Equation (11) uses R^{k',l} with an initial state R^{k',1}, but R^{k',0} is never defined, and the dimensions and initialization of R^{k',1} are not stated. Please clarify how the hierarchical behavior embedding matrix is initialized and updated.","section":"Eq. (11)"},{"comment":"The parameter-analysis figures do not include error bars despite the paper reporting five runs, and some axis labels contain typos such as 'T aobao'. Adding error bars would make the claims about optimal layer counts and scaling factors more convincing.","section":"Figures 3–5"},{"comment":"The description says duplicate user-item interactions are eliminated by retaining the earliest one, but it is unclear whether this rule is applied per behavior type or across all behaviors, and how it affects the construction of B_k. Please clarify the preprocessing for each behavior matrix.","section":"§5.1.2, Table 1"},{"comment":"The module name is introduced as 'Distributed Fitting Multi-Expert Network' in the abstract and Section 4.3, but the conclusion calls it 'Distribution Fitting Expert Network'. Please use one consistent name throughout.","section":"Abstract and §6"}],"recommendation":"major_revision","confidential_remarks":"The missing evaluation split is the single most important issue. I am not accusing the authors of intentional leakage, but the paper as written cannot be distinguished from a label-leakage setup because test positives in the behavior matrices would propagate through Eqs. (5), (6), and (11). Please require a precise protocol description and verify that the released code excludes test edges from all adjacency matrices. The combinatorial-optimization framing also needs to be either formalized or softened, since the current text overstates what is actually established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent applied recsys paper that deserves a serious referee, but the central empirical claim cannot be fully trusted as written because the evaluation protocol is underspecified. The stress-test concern about label leakage in the adjacency matrices is legitimate; the paper never states how the split is constructed or whether target test edges are removed from the B_k matrices used in Eqs. (5) and (6). That is load-bearing given the size of the gains (49.91% on Beibei). I don't think it is fatal—the authors likely followed standard pre-processing—but the onus is on them to say so explicitly.\n\nWhat is actually new: COGCN imposes staged pre-, in-, and post-behavior constraints on multi-behavior fusion, a relaxed alternative to strict cascading. DFME adds contrastive distribution alignment, a behavior-fitting expert, and stop-gradient decoupling in the MTL aggregation. The components are individually known, but the combination is a genuine new system. The ablation study is thorough and shows each module contributes, and the compatibility analysis showing DFME improves other backbones is a useful addition.\n\nSoft spots, in proportion: (1) The evaluation protocol is not specified: no split description, no negative sampling details, no validation split for the hyperparameter search. This is the biggest issue. (2) They ran experiments 5 times but report no standard deviations; the significance star is unbacked by any test description. (3) The combinatorial optimization framing is just narrative; there is no formal objective being optimized. Fine as motivation, but the title and abstract oversell it. (4) Eq. (11) is underspecified: the behavior-fitting expert uses a graph convolution with an extra R matrix whose initial state and update are described loosely. (5) The assumption of a globally shared behavior order is a modeling choice, not an observed fact—Appendix A.2 admits timings are unknown. That is a real limitation, though not a fatal one.\n\nThe math in the GCN part is straightforward and internally consistent; nothing in the derivations set off alarms. But as written, the paper is not reproducible enough to accept the SOTA claim at face value. I would send it to review and ask the authors for a precise data protocol, error bars, and a cleaner spec of the fitting expert.\n\nRecommendation: accept for peer review, with major revisions expected. For a reading group, maybe—the architecture is worth discussing, but the evaluation gap tempers enthusiasm.","headline":"A solid incremental multi-behavior recommendation system with large reported gains, but the evaluation protocol is underspecified enough that the headline numbers need scrutiny before they carry weight.","tokens_in":21724,"tokens_out":2311,"would_cite":true,"duration_ms":23311,"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":"Multi-behavior recommendation improves when fusion is treated as a constrained combinatorial search and prediction decouples feature and label signals.","keywords":["combinatorial optimization","multi-behavior recommendation","multi-task learning","graph convolutional network","negative transfer","contrastive learning","recommender systems"],"falsifier":"On a dataset with timestamps for every user-item behavior event, rebuild the behavior order per user from the actual event times and retrain COPF with the constraint order replaced by the observed order. If COPF's advantage over unconstrained or cascading baselines shrinks or disappears under the true order, then the reported gains are an artifact of the assumed view→cart→buy ordering rather than a general property of staged constraints.","tokens_in":20689,"feed_emoji":"🛒","tokens_out":7011,"duration_ms":59990,"temperature":0.7,"pith_summary":"This paper tries to establish that multi-behavior recommendation should be built as a two-step pipeline with explicit structure: a fusion step that treats user behavior patterns as a combinatorial search space and prunes it with staged constraints, and a prediction step that coordinates target and auxiliary tasks so that feature and label distribution gaps do not cause negative transfer. It proposes COPF, made of COGCN for fusion and DFME for prediction, and reports Hit Ratio@10 and NDCG@10 on Beibei, Taobao, and Tmall, with relative HR improvements of 49.91%, 12.06%, and 24.12% over the best baseline on each dataset. A sympathetic reading of the paper is that its staged constraints—pre-behavior, in-behavior, and post-behavior—capture user behavior patterns better than both unconstrained aggregation and strict cascading paradigms, and that its decoupling and stop-gradient designs make auxiliary tasks useful rather than harmful. If correct, the framework gives a reusable recipe for squeezing signal out of sparse, multi-behavior interaction data.","feed_headline":"Staged constraints lift multi-behavior recommendation on three datasets","feed_subtitle":"COPF prunes behavior-pattern space and decouples task gradients, beating the best prior model on all three datasets.","key_machinery":"The load-bearing mechanism is a staged constraint scheme on user behavior patterns, formalized in Definitions 3–5: a pre-behavior constraint channels only upstream behavior outputs into the current behavior's encoder, an in-behavior constraint restricts graph convolution on behavior k to adjacency matrices of behaviors 1...k (Equation 5), and a post-behavior constraint keeps each behavior's output separate for downstream tasks. The second mechanism is DFME, which coordinates the multi-task prediction head: contrastive loss (Equation 8), behavior-fitting experts built by mixing the current and another behavior's representations with small scaling coefficients (Equation 10), and a stop-gradient operation on the auxiliary-to-target aggregation path (Equation 14). Together they turn fusion into a pruned search and prevent auxiliary tasks from corrupting target-task gradients.","core_discovery":"On its own terms, the paper's central claim is that the two weaknesses of current multi-behavior recommenders—imprecise fusion and uncoordinated multi-task prediction—can be fixed by imposing structure at the right points. COGCN formulates fusion as combinatorial optimization and restricts the solution space with three constraints: pre-behavior (the encoder of behavior k receives only outputs of upstream behaviors), in-behavior (message passing over the current behavior's adjacency matrix never uses semantic information from downstream behaviors), and post-behavior (decoupled outputs for each behavior feed separate prediction tasks). DFME then coordinates prediction: contrastive learning aligns target and auxiliary behavior representations, behavior-fitting experts use small mixing coefficients to refine the target's representation space, and a stop-gradient operation prevents auxiliary-task gradients from updating the target task. The paper argues, with ablations, that removing or relaxing any of these stages reduces performance, and that DFME plugs into other backbones and improves them.","pith_inferences":["Editorial inference: the fixed global order assumption is the point to attack; the Appendix admits interaction timings are unknown, so a dataset with timestamps could re-order behaviors per user and reveal whether the staged constraints help or merely encode a prior.","Editorial inference: a natural extension is to learn a per-user or per-item probability distribution over behavior order instead of a single shared order, which would relax Definition 2.","Editorial inference: the contrastive-alignment plus stop-gradient pattern could be transferred to other multi-task ranking settings, such as multi-scenario prediction, where label distributions also differ across tasks."],"forward_implications":["If COPF's constraints are right, multi-behavior recommenders do not have to choose between free-form aggregation and overly strict cascades; staged constraints are the better middle ground.","Auxiliary behaviors (views, carts, collects) can be used to improve buy prediction without the gradient interference that hurts many multi-task recommenders.","DFME should work as a drop-in prediction module: the compatibility study shows it raises accuracy when attached to prior fusion backbones such as LightGCN, MB-CGCN, CRGCN, CIGF, and PKEF.","The framework's success on three datasets with different behavior distributions suggests the gains are not tied to one platform's interaction patterns."],"supporting_citations":[{"why":"supplies PKEF, the strongest baseline on Beibei and Taobao, whose projection-based aggregation DFME is designed to outperform.","marker":"[31]"},{"why":"supplies BCIPM, the strongest baseline on Tmall, which models target-behavior context before the target.","marker":"[52]"},{"why":"supplies CIGF with its decoupled-input MTL baseline MESI, the starting point DFME extends.","marker":"[15]"},{"why":"supplies MB-CGCN, a cascading graph convolution baseline that imposes strict behavior-order constraints.","marker":"[6]"},{"why":"supplies CRGCN, a cascading residual GCN baseline whose strict cascade the paper contrasts with its staged constraints.","marker":"[51]"},{"why":"supplies MBGCN's heterogeneous-relation GCN aggregation, the basis for COGCN's in-behavior convolution.","marker":"[21]"},{"why":"supplies LightGCN, the plain GCN baseline used to build COGCN and the w/o COGCN ablation.","marker":"[16]"},{"why":"supplies MMOE, a multi-expert MTL baseline compared inside the DFME study.","marker":"[29]"},{"why":"supplies PLE, a multi-expert MTL baseline with task-specific and shared experts, used as comparison.","marker":"[40]"}],"fun_headline_variants":["COPF: combinatorial optimization for fusion and decoupled experts","Behavior fusion as optimization, prediction via gradient decoupling","Three constraints on fusion, target-aware experts beat baselines","COGCN + DFME: new multi-behavior recommendation framework excels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework assumes one fixed, globally shared order of behaviors (for instance, view before cart before buy) and that during fusion it is safe to exclude all downstream behavior information, even though real interaction timings are not observed.","fun_headline_variants_meta":{"raw":{"variants":["COPF: combinatorial optimization for fusion and decoupled experts","Behavior fusion as optimization, prediction via gradient decoupling","Three constraints on fusion, target-aware experts beat baselines","COGCN + DFME: new multi-behavior recommendation framework excels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000212,"raw_usage":{"total_tokens":1449,"prompt_tokens":1004,"completion_tokens":445,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":376}},"tokens_in":620,"tokens_out":445,"duration_ms":5537,"temperature":1.0,"reasoning_tokens":376,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T12:52:27.798693+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a dataset with timestamps for every user-item behavior event, rebuild the behavior order per user from the actual event times and retrain COPF with the constraint order replaced by the observed order. If COPF's advantage over unconstrained or cascading baselines shrinks or disappears under the true order, then the reported gains are an artifact of the assumed view→cart→buy ordering rather than a general property of staged constraints.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies PKEF, the strongest baseline on Beibei and Taobao, whose projection-based aggregation DFME is designed to outperform."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies BCIPM, the strongest baseline on Tmall, which models target-behavior context before the target."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies CIGF with its decoupled-input MTL baseline MESI, the starting point DFME extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies MB-CGCN, a cascading graph convolution baseline that imposes strict behavior-order constraints."},{"cited_title":"Cascading Residual Graph Convolutional Network for Multi-Behavior Recommendation","cited_arxiv_id":"2205.13128","evidence_quote":"supplies CRGCN, a cascading residual GCN baseline whose strict cascade the paper contrasts with its staged constraints."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies MBGCN's heterogeneous-relation GCN aggregation, the basis for COGCN's in-behavior convolution."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies MMOE, a multi-expert MTL baseline compared inside the DFME study."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies PLE, a multi-expert MTL baseline with task-specific and shared experts, used as comparison."}],"review_version":1}