{"id":"bef666ff-13fd-400f-bc6d-f752c82bcc27","arxiv_id":"2412.11747","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"TMLP models multimodal item relations with MLPs plus topological pruning and mutual-information learning, outperforming GCN-based recommenders on three Amazon datasets.","lead":"This paper introduces TMLP, a multimodal recommender that replaces graph convolutional item modeling with topology-aware MLPs. On three Amazon datasets it beats nine baselines on accuracy, but some efficiency claims are overstated.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central attribution to topology-aware MLPs is confounded: the ablations do not rule out large-capacity MLPs plus feature-similarity supervision as the actual source of the gains.","rationale":"Reader's verdict CONDITIONAL is appropriate. I agree with the reader's weakest assumption; it is the same confound I identify. The reason I would not move to REJECT is that the paper has several genuine strengths: the code is public, the ablations include multiple controls (w/o pruning, random pruning, w/o NA loss), stds over 5 seeds are reported, and significance tests are stated. The reason I would not ACCEPT is that the central mechanism is not yet identified: the 'topology-aware' label could be an epiphenomenon of feature-space smoothing plus added MLP capacity. The proposed test directly separates the topological supervision from the feature-similarity signal by replacing the TPS graph with raw-feature neighbors while keeping every other component fixed. If TMLP still beats this control, the concern is resolved; if not, the paper's attribution needs revision, though the empirical method may still be useful. The reader's additional overclaim concerns about the '7%' phrasing and efficiency uniformity are secondary and do not change this assessment.","tokens_in":15320,"tokens_out":6032,"duration_ms":61960,"concrete_test":"Run a new ablation on Sports and Baby with the same TMLP architecture and hyperparameters (Eq. 9, alpha, K, LightGCN aggregator), but replace the TPS-pruned adjacency \\bar A in Eq. 8 with a binary adjacency whose entries are 1 for the top-K raw-feature cosine neighbors (using the same beta_m-weighted features that build A in Eq. 2), bypassing Eqs. 5-6 while keeping the same number of positives, negatives, temperature, and loss weight. If this Feature-Sim variant reaches Recall@20 within 1% relative of TMLP's 0.1152 (Sports) or 0.1016 (Baby), the topology-specific mechanism is not load-bearing; if it drops by more than about 10% relative, the topological claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is causal: replacing GCN message passing with TPS-pruned neighborhood alignment (Eqs. 5-8) is why TMLP beats the nine baselines. The load-bearing assumption is that the injected supervision is topological signal, not just smoothed pretrained-feature similarity, and that the larger MLP encoder (Eq. 9, hidden 512, 2-4 layers) is not itself the source of the gains. Table 3's ablations show NA loss and TPS matter, but they leave this confound open: 'w/o Pruning' still supervises with a graph built from the same pretrained features, and 'Rand Pruning' preserves the same average degree and the same feature-derived edge set, so neither distinguishes 'topology matters' from 'pulling together similar raw features helps.' The appendix reports small stds, so the empirical gap is real, but it is not evidence for the topology-specific mechanism. Without a same-capacity control that replaces graph supervision with raw-feature nearest-neighbor supervision, the central claim is not settled.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TMLP, a multimodal recommender system that replaces GCN-based item-item relation modeling with MLPs. Two auxiliary components are introduced: a Topological Pruning Strategy (TPS) that denoises an item-item graph built from pretrained visual and textual features, and an intra/inter-modality learning objective based on a neighborhood-alignment (NA) mutual-information loss. Experiments on Amazon Baby, Sports, and Electronics report consistent improvements over nine baselines in Recall and NDCG, accompanied by ablations, hyperparameter sensitivity studies, training-efficiency comparisons, and a robustness test under corrupted item-item graphs.","tokens_in":15553,"tokens_out":5269,"duration_ms":49624,"significance":"The central idea is interesting and practically relevant: if MLPs with auxiliary topological supervision can match or exceed shallow GCNs in multimodal recommendation, this would avoid over-smoothing and improve scalability. The paper has concrete strengths: multi-seed experiments with t-tests and standard deviations in the ablations, a robustness check, a case study, and a code release link. The empirical gains are consistently positive across datasets. However, the mechanism attribution is not fully settled, and one efficiency claim is contradicted by the paper's own runtime table, so the headline conclusions need revision and additional experiments.","major_comments":[{"comment":"The abstract and contributions claim 'achieving over a 7% performance gain on Amazon Baby'. In Table 2, the improvement over the second-best model on Baby is 7.53% for Recall@10, but only 3.15% for R@20, 6.19% for N@10, and 5.15% for N@20; on Sports and Electronics the gains range from 3.83% to 6.84%. The headline claim is therefore metric-selective and should be qualified as 'up to 7.53% on Recall@10' or presented as a range across metrics.","section":"§5.2 (Table 2) and Abstract"},{"comment":"The paper repeatedly states that TMLP achieves 'significant improvements in training efficiency' (Abstract and §5.5). However, Table 6 reports a total training cost of 10.2 minutes for TMLP versus 8.9 minutes for MGCN, with MGCN reaching R@20 = 0.1106 in 64 epochs and TMLP in 73 epochs. Thus the efficiency advantage does not hold for MGCN in total wall-clock time. The claim should be scoped to per-epoch time or to the other baselines, or the total-cost comparison should be revisited after controlling for convergence criteria and hardware.","section":"§5.5 and Table 6"},{"comment":"The ablations do not settle the paper's central attribution that topological structure, as captured by TPS and the neighborhood-alignment loss, drives the gains. 'w/o Pruning' still supervises with the item-item graph A built from the same pretrained visual/textual features, and 'Rand Pruning' keeps the same degree and the same feature-derived edge set. Neither variant replaces graph supervision with direct raw-feature nearest-neighbor supervision (e.g., kNN on the raw visual or textual embeddings without constructing A) or increases MLP capacity to match the removed supervision. Without such a same-capacity control, the observed improvements could come from pulling together semantically similar pretrained features rather than from topology awareness. Please add this control; if it performs similarly, the 'topology-aware' mechanism claim should be weakened.","section":"§5.3 (Table 3), Eqs. (2)–(8)"},{"comment":"The paper does not document hyperparameter tuning for the nine baselines. It states that TMLP's hyperparameters are searched over the grids listed in §5.1, but for baselines it only refers to MMRec and 'same settings as Zhou et al. (2023)'. If the baselines were not individually tuned on each dataset, the SOTA comparisons may be unfair. I ask the authors to report the search ranges and selected hyperparameters for each baseline, or to state explicitly that they were taken verbatim from prior work without tuning, and to confirm that the same evaluation split and ranking protocol were applied.","section":"§5.1 Implementation Details"}],"minor_comments":[{"comment":"The phrase 'discarding the internal message passing in GCNs' is imprecise because TMLP retains LightGCN layers on the user-item interaction graph (Eqs. 10–12); the contribution concerns item-item relation modeling specifically.","section":"Abstract and Introduction"},{"comment":"There are several typos, including 'lrocessing' in the Introduction, 'Univeristy' in the affiliation, 'Objctive' in §4.5, and 'dependices' in Eqs. (5)–(6) and in the Training and Inference paragraph. A careful proofread is needed.","section":"§1 and throughout"},{"comment":"The notation 1_{n≠m}(Āmn)^r in the numerator is confusing because the same symbol Āmn acts as both a mask and an exponent, and only r=1 is used in the experiments. Please clarify the general definition or state r=1 directly in the equation.","section":"Eq. (8)"},{"comment":"Table 2 reports only means and starred significance, while the appendix reports standard deviations only for ablations. Please add standard deviations to Table 2 as well, since the variability across the five seeds is relevant for the claimed margins over the second-best baseline.","section":"Table 2 and Table 5"},{"comment":"The case study is qualitative. It would be more convincing to quantify the effect of TPS, for example by measuring the fraction of pruned edges that connect items with different ground-truth category labels, or by reporting a small human-evaluation agreement score.","section":"§5.7"}],"recommendation":"major_revision","confidential_remarks":"The empirical work is solid in execution, but the paper's own Table 6 undermines the general training-efficiency claim for one of the strongest baselines, and the ablation design does not yet isolate topology awareness from raw-feature-similarity supervision. Both are load-bearing for the abstract's central claims. I believe these can be fixed with additional experiments and scoped wording, hence major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a solid empirical paper worth engaging, but the causal claim about topology-awareness is stronger than the evidence supports. The authors replace GCN message passing with MLPs for multimodal item-item relations, using a topological pruning step and a neighborhood-alignment mutual information loss. On three Amazon datasets with five seeds and significance tests, TMLP beats nine baselines. The gains are real and the combination is new: prior MMRS work uses GCNs for item-item modeling; this is the first to do it with MLPs, borrowing Graph-MLP and N2N ideas. The paper is honest about components and releases code.\n\nSoft spots. The headline 'over 7%' is metric-selective: it is Recall@10 on Baby; other metrics are 3-6%. The efficiency claim is not uniform: Table 6 shows MGCN trains in 8.9 minutes vs TMLP's 10.2 on Sports, even though TMLP converges in fewer epochs. The paper says TMLP has shorter time per epoch and faster convergence, but total training cost is not better than MGCN. Bigger issue: the ablation does not isolate topology. The NA loss aligns a node with neighbors from a graph built on pretrained feature similarity. A high-capacity MLP (hidden 512) with feature-similarity supervision could produce the same gains. 'w/o Pruning' still supervises with the same feature-derived graph; 'Rand Pruning' preserves average degree and the same edge set. Neither rules out 'similar features help' as the driver. A control with a same-capacity MLP trained on the raw kNN graph, without TPS, would settle it. Baseline tuning is also under-documented.\n\nWho this is for: multimodal recommender people and anyone working on graph-free GNN alternatives. It deserves a serious referee; I would accept it for review with major-revision expectations, not desk reject. The empirical core is reproducible and the gains are not artifacts, but the mechanism claim and the efficiency claim need tightening.","headline":"A genuinely useful empirical paper on MLP-based multimodal recommendation, but the causal claim about topology is not fully pinned down and the efficiency story is overstated.","tokens_in":16095,"tokens_out":2413,"would_cite":true,"duration_ms":23225,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper establishes that a topology-aware MLP, trained with a neighborhood-alignment mutual-information loss on a topologically pruned item graph, beats nine baselines on three Amazon datasets, improving Recall@10 by 7.53%, 5.49%, and…","keywords":["multimodal recommendation","graph convolutional networks","MLP","topological pruning","mutual information","contrastive learning","over-smoothing","collaborative filtering"],"falsifier":"Measure TMLP against its own w/o Pruning variant on a dataset whose visual and textual features are randomly permuted across items before building the item-item graph; if TMLP's margin over w/o Pruning persists under permutation, the improvement is not attributable to topological denoising of real semantics.","tokens_in":15116,"feed_emoji":"📈","tokens_out":8092,"duration_ms":62634,"temperature":0.7,"pith_summary":"This paper claims that graph convolution is not needed to capture item-item topology in multimodal recommendation. It proposes TMLP, which replaces GCN message passing with MLPs and teaches the MLPs topology through two complementary mechanisms: a topological pruning strategy that denoises the item-item graph built from pretrained visual and textual features, and a neighborhood-alignment mutual-information loss that pulls representations of topologically related items together. On the Amazon Baby, Sports, and Electronics benchmarks, TMLP reports Recall@10 improvements of 7.53%, 5.49%, and 5.69% over the best baseline, along with faster convergence and better robustness to corrupted item connections. If correct, this suggests over-smoothing is not an inherent tax on deep multimodal recommenders, and that MLP-based models can scale past shallow GCNs.","feed_headline":"Topology-aware MLPs beat GCN recommenders by up to 7.5%","feed_subtitle":"Replacing message passing with pruned-graph mutual information lifts Recall@10 on three Amazon benchmarks.","key_machinery":"The central mechanism is replacing graph convolution with a mutual-information alignment objective over a pruned item graph. The Topological Pruning Strategy (TPS) computes, for each item, the topological similarity $$TS(X_m;X_n)=\\sum_{x_m}\\sum_{x_n} p(x_m,x_n)\\log\\frac{p(x_m,x_n)}{p(x_m)p(x_n)}$$ between its neighborhood-membership distribution and each neighbor's, and retains the top-K edges per node; this runs once before training. The Intra (Inter)-Modality Learning (IML) module then injects this topology into an MLP: per-modality MLP encoders produce visual and textual item representations, a fuser combines them, and the neighborhood-alignment (NA) loss $$L_{NA} = -\\mathbb{E}_{v_m}\\left[\\log \\frac{\\sum_{n} \\mathbb{1}[\\bar{A}_{mn}]^r \\exp(\\operatorname{sim}(z_m,z_n)/\\tau)}{\\sum_{k \\neq m} \\exp(\\operatorname{sim}(z_m,z_k)/\\tau)}\\right]$$ maximizes mutual information between an item and its r-hop neighbors under MINE estimation. The joint objective $L = L_{BPR} + \\alpha L_{NA} + \\lambda\\|\\Theta\\|^2$ makes the MLP topology-aware without any message passing at inference.","core_discovery":"TMLP establishes that the performance gains previously attributed to GCN message passing in multimodal recommendation can be obtained, and surpassed, by a topology-aware MLP. The item-item graph is first denoised by TPS, which keeps only the top-K neighbours per item according to a topological similarity score computed as the mutual information between neighborhood membership distributions (Equation 5). The MLP is then trained with the BPR ranking loss plus a neighborhood-alignment (NA) contrastive loss (Equation 8) that maximizes mutual information between an item's representation and its r-hop neighbours, implemented through MINE; an optional fuser combines visual and textual MLP branches. The resulting item representations feed a LightGCN aggregator on the user-item graph for final ranking. Empirically, TMLP outperforms nine baselines on three Amazon datasets, improves Recall@10 by 7.53%, 5.49%, and 5.69% over the second-best model, converges in fewer epochs, and maintains performance when up to 10% of item-item edges are corrupted.","pith_inferences":["A natural extension is to learn the pruning decision per dataset instead of fixing K: random pruning was competitive on Sports at K=5, so a learned edge-dropping rule might improve TMLP further on datasets with different degree distributions.","Because the NA loss is a contrastive objective on a denoised graph, the same TPS-plus-MI recipe could transfer to other graph-based prediction tasks that currently rely on shallow GCNs, such as node classification on small graphs.","If the reported gains largely stem from MLP capacity rather than topology awareness, then a plain MLP with the same width, depth, and BPR loss but no NA loss should close much of the gap; the paper's w/o NA Loss ablation suggests it does not, but that comparison is only shown on two datasets.","The paper's claim that unrelated items stay separated under depth (Figure 1d) implies a testable diagnostic: on larger catalogs, TMLP's item embeddings should retain higher average pairwise cosine distance than GCN baselines at every depth."],"forward_implications":["Performance on the Amazon Electronics dataset keeps rising with MLP depth, while GCN-based baselines degrade after one layer, indicating TMLP sidesteps over-smoothing rather than merely tolerating it.","Because topological pruning is done once before training and MLP forward passes do not propagate messages, TMLP reaches peak Recall@20 in roughly 10 to 30 epochs, compared with about 200 for FREEDOM and BM3, and its per-epoch cost is lower.","When the item-item graph is corrupted at 1% and 10% noise ratios, TMLP's Recall@20 on Sports stays nearly flat, while FREEDOM and MGCN drop substantially, so the pruning step appears to make the model robust to modality mismatch.","At inference the model needs only item features, not the item-item graph, so the learned representations can be served without maintaining graph topology."],"supporting_citations":[{"why":"Supplies the method of constructing modality-specific item similarity graphs (kNN on pretrained features) and their weighted fusion used to build the item-item graph A.","marker":"(Zhang et al. 2021a)"},{"why":"FREEDOM is the strongest baseline and provides the item-item graph construction and freezing approach that TMLP's TPS prunes and denoises.","marker":"(Zhou and Shen 2023)"},{"why":"LightGCN is the CF aggregator TMLP uses on the user-item graph and one of the general-CF baselines.","marker":"(He et al. 2020)"},{"why":"MINE is the mutual-information neural estimation technique that converts the NA objective into a tractable contrastive loss.","marker":"(Belghazi et al. 2018)"},{"why":"Provides the node-to-neighbourhood mutual information maximization idea that motivates both the NA loss and the topological similarity score.","marker":"(Dong et al. 2022)"},{"why":"MGCN is a state-of-the-art GCN baseline whose over-smoothing and attention-based modality fusion TMLP is compared against.","marker":"(Yu et al. 2023)"},{"why":"BM3 is a baseline and the source of the pretrained 4096-d visual and 384-d textual features and the 8:1:1 data split.","marker":"(Zhou et al. 2023)"},{"why":"BPR provides the ranking loss used in the joint objective and is the matrix-factorization baseline.","marker":"(Rendle et al. 2012)"}],"fun_headline_variants":["Topology-aware MLPs outperform GCNs in multimodal recommender systems","MLPs beat GCNs in multimodal recommendation via topology-aware pruning","TMLP: MLP-based multimodal recommender surpasses GCN baselines","Pruning and mutual information let MLPs beat GCNs in recommendation","Drop GCNs: topology-aware MLPs keep multimodal recommendation strong"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the item-item graph built from pretrained visual and textual features carries genuine semantic similarity after pruning, rather than encoding the same modality mismatch the pruning is meant to remove; if that graph is mostly noise, the gains could come from extra MLP capacity instead of topology awareness.","fun_headline_variants_meta":{"raw":{"variants":["Topology-aware MLPs outperform GCNs in multimodal recommender systems","MLPs beat GCNs in multimodal recommendation via topology-aware pruning","TMLP: MLP-based multimodal recommender surpasses GCN baselines","Pruning and mutual information let MLPs beat GCNs in recommendation","Drop GCNs: topology-aware MLPs keep multimodal recommendation strong"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001228,"raw_usage":{"total_tokens":5072,"prompt_tokens":998,"completion_tokens":4074,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":614,"completion_tokens_details":{"reasoning_tokens":3978}},"tokens_in":614,"tokens_out":4074,"duration_ms":25936,"temperature":1.0,"reasoning_tokens":3978,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:37:22.151301+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure TMLP against its own w/o Pruning variant on a dataset whose visual and textual features are randomly permuted across items before building the item-item graph; if TMLP's margin over w/o Pruning persists under permutation, the improvement is not attributable to topological denoising of real semantics.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the node-to-neighbourhood mutual information maximization idea that motivates both the NA loss and the topological similarity score."}],"review_version":1}