{"id":"2a3171b3-2c6f-4fa0-a21e-5c4b198a70ff","arxiv_id":"2501.13908","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CDE-CF, a graph neural ODE recommender whose convolution weight is generated by an MLP at each time step, reports slightly better Recall@20 and NDCG@20 than GODE-CF and other baselines on four Amazon datasets.","lead":"This paper swaps the fixed weight in a graph-based neural ODE recommender for a learned weight that changes over time and across users, generated by a small neural network. The authors report small but consistent ranking gains over the strongest ODE-based baseline on four Amazon review datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 reports single-run scores with no variance or significance tests; the smallest CDE-vs-GODE gap is 0.06% relative, so 'consistently outperforms' is not yet established.","rationale":"Reader identified the lack of error bars and significance tests as the weakest assumption, and I agree. It is load-bearing because the whole contribution is the empirical improvement; the mathematical framing aside, Table 1 is the only evidence for the headline claim. The method is simple and plausible, and the code is available, so a repeated-run test is cheap and decisive. I would keep the CONDITIONAL verdict: the mechanism (continuous gating via MLP inside the ODE) is clearly described in Eq. 3 and the improvements are directionally consistent, but the magnitudes are within the range where seed-level noise typically matters. Secondary concerns that do not change the verdict: (i) calling Eq. 3 a 'controlled differential equation' is imprecise—there is no external control path X(t), so it is an ODE with state-dependent gating; this affects framing, not the empirical comparison. (ii) Section 3 text says the MLP input is the initial embedding while Eq. 3 uses E_u(t), and the released code should be checked to confirm which is implemented; if only initial embeddings are used, the 'continuous over time' claim is wrong. Neither is as decisive as the missing variance, and both can be resolved from the public code.","tokens_in":7258,"tokens_out":7828,"duration_ms":70917,"concrete_test":"Run CDE-CF, GODE-CF, and the leading GCN baselines (LightGCN, GTN) for 10 random seeds each on all four datasets using the released code and reported hyperparameters; report mean ± std of Recall@20 and NDCG@20 and apply a paired significance test (e.g., paired t-test or Wilcoxon signed-rank) comparing CDE-CF to GODE-CF per dataset. If the CDE-CF advantage is not significant at p<0.05 on at least three of the four datasets, the consistent-superiority claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is the consistent superiority of CDE-CF over GODE-CF and GCN baselines (Abstract; Section 4.1, Table 1). The evidence consists of one point estimate per method per dataset. No standard deviation, confidence interval, or significance test is reported anywhere. The margins over the closest ODE baseline are tiny: Health Recall +2.4% relative (0.03468 vs 0.03387), Cell Phone Recall +0.06% relative (0.05082 vs 0.05079), and several NDCG gaps below 0.5% relative. The ablation Table 3 has the same single-run character (e.g., Cell Phone Recall 0.05079 vs 0.05082). Under leave-one-out full ranking, run-to-run variation from initialization, solver tolerance, and data-order can easily exceed these gaps. If repeated runs show the difference is within noise on any dataset, the abstract's 'surpasses' and 'consistently outperforms' claims would not be supported, leaving only a marginal direction. The code is public, so the check is straightforward.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CDE-CF, a collaborative filtering method that augments the GODE-CF framework by replacing its discrete weight matrix with a continuous weight generated by an MLP with a sigmoid activation, integrated as part of the Neural ODE. The model is trained with BPR loss and evaluated with Recall@20 and NDCG@20 on four Amazon datasets. The authors report that CDE-CF consistently outperforms GCN-based and ODE-based baselines, and they provide an ablation study on the weight component and a training-time comparison.","tokens_in":7530,"tokens_out":4362,"duration_ms":41256,"significance":"If the empirical claims hold, the contribution is a modest but useful modification of an existing ODE-based CF method: it replaces a discrete weight with a continuous, MLP-generated weight and shows that this can improve ranking accuracy. The paper is honest in scope, includes a public code repository, and contains an ablation that directly tests the continuous-weight design. The main weakness is that the central superiority claim is supported only by single-run point estimates with no uncertainty quantification; several of the reported gaps are tiny, so the statistical significance of the improvement is not yet established.","major_comments":[{"comment":"The central claim that CDE-CF consistently outperforms GODE-CF rests on a single point estimate per method per dataset, with no standard deviations, confidence intervals, or significance tests. The margins are very small in several cases: Cell Phone Recall is 0.05082 vs 0.05079 (0.06% relative) and Beauty NDCG is 0.03426 vs 0.03406 (0.59% relative). Under full-ranking evaluation, such differences can easily arise from random seed, solver tolerance, or data-order variation. Please report means and standard deviations over multiple runs and provide a statistical significance test (e.g., paired tests over seeds) for the CDE-CF versus GODE-CF comparison.","section":"Section 4.1, Table 1"},{"comment":"The training-time comparison is confounded by method-specific integration horizons: the text states that the optimal t is about 8.5 for GODE-CF and about 6.5 for CDE-CF. Since the number of solver steps and hence training time scale with t, the speed comparison is not a like-for-like test of the methods. Please report the exact t used for every method, describe how t was selected (validation or test), and include a sensitivity analysis over t to show that the performance and efficiency conclusions are robust.","section":"Section 5, Table 4"},{"comment":"The formulation in Eq. (3) is underspecified and internally inconsistent. The preceding text says the MLP takes the initial embedding as input, but Eq. (3) uses f(E_u(t), theta) and f(E_i(t), theta), i.e., the current-time embedding. In addition, sigma(f(...)) is never defined as a diagonal matrix, so the product sigma(f(E_u(t),theta)) (A_n - I) E_i(t) is ambiguous. Please define the exact tensor dimensions and the precise forward computation, and reconcile the text with the equation.","section":"Section 3, Eq. (3)"}],"minor_comments":[{"comment":"The second and third contribution bullets contain naming errors: 'ODE-CF' should be 'CDE-CF', and the phrase 'same training time as CDE-CF' is self-referential and should be corrected (presumably to 'as GODE-CF' or 'as the baselines').","section":"Section 1, contributions"},{"comment":"There are typos such as 'Differnt' and 'the the same timestep'; please proofread the manuscript.","section":"Section 3, first paragraph"},{"comment":"The statement that without weights all node embeddings would converge to the same state at the same timestep is asserted without proof or conditions on A_n. Since this motivates the weight-control design, please provide a short justification or explicitly state the assumptions under which it holds.","section":"Section 3, paragraph after Eq. (2)"},{"comment":"The table formatting is inconsistent: the Office NDCG column shows '0.022137' with five decimals while all other entries have five decimals or fewer; please standardize the precision and align the table.","section":"Section 4.1, Table 1"},{"comment":"The paper does not report the optimizer, learning rate, embedding dimension, batch size, or the number of epochs used for the main results; please add these reproducibility details.","section":"Section 4, experimental setup"},{"comment":"The figure appears to be missing or not embedded in the submitted text; if it is meant to include the subplot comparing weight values, please ensure it is actually present and legible.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of a recommender-systems or graph-learning venue, but the contribution is incremental and the empirical evidence is currently too weak to support the abstract's 'consistently outperforms' claim. The main fix is straightforward: multiple seeded runs with significance testing, a clear specification of the model, and a fair comparison of integration horizons. I would also ask the editor to ensure that the same-group baseline GODE-CF is configured as fairly as possible, since the authors share affiliation with that paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a modest but real increment over GODE-CF. The authors swap a discrete weight for an MLP-sigmoid-produced continuous weight inside the graph ODE, and they show a consistent, if small, improvement on four Amazon datasets. The paper is honest about the fact that the discrete weight does not always help. The code is public and the ablation in Table 3 is the right experiment.\n\nThe soft spot is the one the stress test flagged: every number in Table 1 and Table 3 is a single run. No standard deviations, no significance tests. The margins are tiny; the smallest gap is 0.06% relative. With that evidence, \"consistently outperforms\" is a statement of direction, not a result. The paper also tunes the integration horizon separately per method, and it does not report the baseline protocol, so part of the gap could be a tuning artifact.\n\nTwo smaller issues. The title and method name invoke neural CDEs, but the paper never cites the neural CDE literature (Kidger et al., for example), and the model is an ODE with a gated state-dependent weight, not a CDE driven by a control path. The naming will confuse people. Second, experimental details are thin: no seed count, solver tolerance, or MLP architecture, which will slow reproduction.\n\nNone of this kills the paper. The core idea is simple and the evidence, while weak, points in a consistent direction. The ablation supports the continuous-weight benefit directly. Since the code is public, a referee can quickly check whether the margins survive repeated runs. I would send this to peer review with a clear request for repeated runs, significance tests, and a fuller comparison with the neural CDE literature. The paper is not ready for publication as written, but it is a legitimate workshop-or-conference paper with a checkable claim.","headline":"A small, reproducible extension of GODE-CF whose central claim of consistent superiority is undercut by single-run results and tiny margins.","tokens_in":8024,"tokens_out":2838,"would_cite":false,"duration_ms":26502,"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 claims that replacing the discrete weight in the graph-ODE model GODE-CF with a continuous sigmoid-MLP-generated weight yields the highest Recall@20 and NDCG@20 on four Amazon benchmarks, and that continuous weight control is…","keywords":["collaborative filtering","graph neural ordinary differential equations","continuous weight control","graph convolution","recommendation","CDE-CF","BPR loss","Amazon benchmarks"],"falsifier":"Re-run every model on the same four datasets with multiple random seeds and report means with standard deviations or significance tests; if the Recall@20 and NDCG@20 differences between CDE-CF and GODE-CF become indistinguishable from noise, the continuous-weight claim is refuted.","tokens_in":7098,"feed_emoji":"📈","tokens_out":7575,"duration_ms":60302,"temperature":0.7,"pith_summary":"The paper argues that in graph-neural ODE models for collaborative filtering, the weight matrix that controls graph convolution must vary continuously in time and across nodes, not sit fixed. It introduces Graph Neural Controlled Differential Equations for Collaborative Filtering (CDE-CF), which inserts an MLP followed by a sigmoid into the ODE integrand so the weight at each solver step is generated from the current embedding. On the four Amazon benchmarks used, CDE-CF reports the highest Recall@20 and NDCG@20 among all baselines, and the ablation shows it beats both the no-weight and discrete-weight variants in every dataset.","feed_headline":"Continuous weight control lifts graph-ODE recommenders to top scores","feed_subtitle":"Replacing GODE-CF's fixed weight with a sigmoid-MLP controller raises Recall@20 and NDCG@20 on four Amazon datasets.","key_machinery":"The load-bearing object is the continuous weight controller inside the graph-ODE integral: $$E_u^*=E_u(0)+\\$int_0^{{t_1}}$ \\$\\sigma$(f(E_u(t),\\$\\theta$))(A_n-I)E_i(t)\\,dt,$$ and symmetrically for items. The MLP $f$ consumes the current embedding and the sigmoid bounds its output, so the convolution weight adapts to both node identity and integration time; this is the mechanism that distinguishes CDE-CF from GODE-CF's fixed $W$ and from LightGCN's weight-free convolution.","core_discovery":"CDE-CF builds directly on GODE-CF but changes how the convolution weight is produced. Where GODE-CF uses a fixed discrete weight matrix $W$, CDE-CF replaces it inside the ODE with $\\sigma(f(E_u(t),\\theta))$ and $\\sigma(f(E_i(t),\\theta))$, where $f$ is an MLP applied to the current user or item embedding at time $t$ and $\\sigma$ is the sigmoid. The weight therefore changes along each node's own trajectory, and the paper presents this time- and node-dependent control as the reason CDE-CF outperforms fixed-weight, no-weight, and GCN baselines on all four datasets.","pith_inferences":["If the gains replicate under multiple seeds, the same MLP-plus-sigmoid controller could be transplanted to other graph-neural ODE applications, such as node classification or temporal graph models, where nodes also need different integration speeds.","A natural extension the paper leaves open is conditioning the weight generator on edge features or global context rather than only the current embedding, which might widen the margin.","A testable prediction is that denser graphs need a smaller optimal $t_1$; sweeping $t_1$ on sparsity variants of the four datasets would check whether the continuous controller changes how quickly the ODE should stop.","A direct comparison with LT-OCF's learnable-time mechanism would clarify whether continuous weight control or learned integration time is the more effective form of adaptability in graph-ODE recommenders."],"forward_implications":["If the reported results hold, continuous weight control is a general improvement over discrete or absent weights for ODE-based collaborative filtering.","The optimal integration time $t_1$ drops from about 8.5 for GODE-CF to about 6.5 for CDE-CF, so the same accuracy is reached with fewer ODE solver steps.","The ablation on all four datasets shows continuous weights beat both no-weight and discrete-weight variants, indicating the benefit comes from time-adaptive weighting rather than from having a weight at all.","Training time stays competitive with GODE-CF and is faster on three of the four datasets despite the extra MLP component."],"supporting_citations":[{"why":"GODE-CF is the baseline that CDE-CF extends; it supplies the ODE formulation with a discrete weight matrix and the main comparison point.","marker":"[12]"},{"why":"LightGCN is the simplified linear GCN baseline and the design principle that ODE-based collaborative filtering methods build on.","marker":"[6]"},{"why":"LT-OCF is the learnable-time neural-ODE collaborative filtering baseline that CDE-CF outperforms.","marker":"[2]"},{"why":"Graph Neural Ordinary Differential Equations provides the graph-ODE basis for parameterizing the derivative with GCN layers.","marker":"[10]"},{"why":"The Amazon Reviews dataset is the source of the four benchmark categories used for evaluation.","marker":"[9]"},{"why":"BPR loss is the ranking objective used to train the embeddings in CDE-CF and the baselines.","marker":"[11]"},{"why":"Neural Ordinary Differential Equations supplies the general NODE framework and solver background behind the integral formulation.","marker":"[1]"}],"fun_headline_variants":["Adaptive weight control in graph ODEs boosts recommendation accuracy","MLP-driven continuous weights improve graph ODE recommenders","Time-varying weight control sharpens ODE-based collaborative filtering","CDE-CF: Continuous weight control lifts graph-ODE recommenders"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that CDE-CF beats GODE-CF rests on single-run point estimates in Table 1; the smallest gap (Cell Phone Recall@20) is 0.06% relative, so if repeated runs put the margins within noise, the central conclusion collapses.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive weight control in graph ODEs boosts recommendation accuracy","MLP-driven continuous weights improve graph ODE recommenders","Time-varying weight control sharpens ODE-based collaborative filtering","CDE-CF: Continuous weight control lifts graph-ODE recommenders"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001078,"raw_usage":{"total_tokens":4483,"prompt_tokens":893,"completion_tokens":3590,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":509,"completion_tokens_details":{"reasoning_tokens":3519}},"tokens_in":509,"tokens_out":3590,"duration_ms":23814,"temperature":1.0,"reasoning_tokens":3519,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:28:47.479532+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run every model on the same four datasets with multiple random seeds and report means with standard deviations or significance tests; if the Recall@20 and NDCG@20 differences between CDE-CF and GODE-CF become indistinguishable from noise, the continuous-weight claim is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"LT-OCF is the learnable-time neural-ODE collaborative filtering baseline that CDE-CF outperforms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"LightGCN is the simplified linear GCN baseline and the design principle that ODE-based collaborative filtering methods build on."}],"review_version":1}