{"id":"15553a29-1dc4-42ad-947b-cde0a52584b2","arxiv_id":"2412.14978","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SMORE improves multimodal recommendation by denoising and fusing visual and textual features with FFT-based spectral filters and graph learning.","lead":"A new recommender system fuses product images and text in the frequency domain, filtering out noisy components before combining them, and reports consistent accuracy gains on three Amazon datasets. This matters because real shopping data often contains blurry images and irrelevant descriptions that can distort recommendations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (5) multiplies the unfiltered spectra eH_i,m rather than the filtered bH_i,m from Eq. (4), so as written the fusion path bypasses the modality filter that the central denoising claim rests on.","rationale":"The reader's weakest assumption was that the trainable complex filter in Eq. (4) actually separates signal from modality-specific noise. My review identifies a more concrete and, I argue, more load-bearing problem: even if the filter did denoise, Eq. (5) as written does not use its output in the fusion step. The fusion spectrum is formed by multiplying the original spectra eH_i,m, so the per-modality filter is bypassed exactly where the paper claims noise suppression happens. This is an internal inconsistency in the method description, not merely missing quantitative evidence. Because the central claim is that the model's improvement arises from spectrum-based denoising, this inconsistency directly undermines the explanation of the results. It also creates ambiguity about what was actually implemented and evaluated: if the code uses bH in the product, the paper has a typo; if the code uses eH, the claimed mechanism is not operative. In either case, the paper as published cannot be assessed as a sound contribution without resolving this point. Therefore the conditional verdict is appropriate: acceptance should require the authors to confirm which variant is correct, correct the equation if needed, and show that the corrected mechanism reproduces the reported gains. I do not reject the paper outright because the inconsistency is easily fixable and the code is available, but the current text leaves the central claim's mechanism unverified.","tokens_in":14893,"tokens_out":3940,"duration_ms":24170,"concrete_test":"Inspect the released code (github.com/kennethorq/SMORE) for the fusion operation corresponding to Eq. (5): check whether the pointwise product before δ_f uses the original FFT output eH or the filtered bH from Eq. (4). Then, if the code uses bH, rerun the reported experiments with Eq. (5) corrected to use bH and confirm Table 2 results are reproduced; if the code uses eH, replace eH with bH in Eq. (5) and measure the change in Recall@20 on Baby, Sports, and Clothing. A material change would show the published equation does not match the evaluated model, and the claimed denoising-before-fusion mechanism is not what was tested.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central mechanism is that spectral filtering suppresses modality noise before fusion (Section 4.1). Eq. (4) applies a trainable complex filter to each modality's spectrum: bH_i,m = δ_m(eH_i,m). However, Eq. (5) defines the fusion spectrum as bH_i,f = δ_f( Π_{m∈M} eH_i,m ), using the original spectra eH_i,m, not the filtered spectra bH_i,m. Thus, as written, the pointwise product that forms the fused spectrum operates on the unfiltered (noisy) modality spectra, so the per-modality denoising of Eq. (4) is bypassed in the fusion path. Only the post-hoc fusion filter δ_f is applied. This contradicts the stated claim that cross-modality noise is suppressed before or during fusion. If Eq. (5) is a typo and should read Π bH_i,m, the paper needs correction and the experimental implementation must be checked to confirm which variant was actually evaluated. If Eq. (5) is intentional, the denoising story is unsupported: the fusion input is exactly the noisy spectrum whose amplification the paper motivates in Fig. 1. Either way, the central claim's mechanism is not established by the text as written. Additionally, the learned filter W^c has no explicit denoising objective, so downstream accuracy alone cannot distinguish noise suppression from added model capacity.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes SMORE, a multimodal recommendation model that projects item modality features into the frequency domain, applies learned complex filters to suppress modality-specific noise, fuses the filtered spectra, and combines the result with graph-based collaborative and modality-aware preference signals. The model is evaluated on three Amazon subsets (Baby, Sports, Clothing) against general and multimodal baselines, reporting consistent improvements in Recall@K and NDCG@K. The central claim is that spectrum-based fusion simultaneously captures uni-modal and fusion preferences while suppressing cross-modality noise.","tokens_in":15212,"tokens_out":4390,"duration_ms":28456,"significance":"If the mechanism works as claimed, the paper makes a useful contribution: frequency-domain fusion is relatively underexplored in multimodal recommendation, the model is largely implementable, the code is publicly released, and the experiments cover three standard datasets with multiple baselines and five seeds. The ablations in Section 5.3 and the hyperparameter analysis in Section 5.4 are informative. However, the manuscript's own equations contain a load-bearing inconsistency in the fusion path, and the denoising claim is supported only by qualitative t-SNE visualizations and indirect downstream accuracy. The significance of the contribution depends on whether the implementation actually follows the intended filtered-fusion variant.","major_comments":[{"comment":"As written, the fusion spectrum is computed as bH_i,f = δ_f(∏_{m∈M} eH_i,m), i.e., the pointwise product is taken over the unfiltered spectra eH_i,m from Eq. (2), not over the filtered spectra bH_i,m from Eq. (4). This means the per-modality denoising filter δ_m is bypassed in the fusion path, and the sentence following Eq. (5) claiming that the pointwise product captures correlations 'while minimizing noise contamination during fusion' is not supported by the equations. If the intended operation is ∏ bH_i,m, the equation and the implementation must be corrected and verified; if Eq. (5) is deliberate, the central denoising-while-fusion claim needs to be reformulated, because the fusion input is exactly the noisy spectrum whose amplification the paper motivates in Fig. 1.","section":"§4.1, Eq. (5)"},{"comment":"RQ4's evidence for denoising is qualitative: the manuscript shows t-SNE plots and kernel density estimates, but provides no quantitative measurement of noise suppression. In addition, the trainable filter W^c in Eq. (4) has no explicit denoising objective; it is trained end-to-end with BPR and contrastive losses, so improved downstream accuracy alone cannot distinguish noise suppression from added model capacity. Please report a quantitative comparison, e.g., an ablation that runs the same model with and without the per-modality filter in the fusion path, and, if feasible, a spectral or representation-level noise metric.","section":"§5.5, Figs. 6 and 7"},{"comment":"The manuscript states that results are stable across five seeds and statistically significant with p < 0.01, but Table 2 reports only mean values with no standard deviations or confidence intervals. This is particularly important because some improvements are small, e.g., Recall@10 on Clothing is 0.0659 for SMORE versus 0.0649 for MGCN and 0.0635 for FREEDOM. Without per-seed variance, the consistency claim cannot be assessed by the reader; please include standard deviations or per-seed results, or justify why the paired t-test is sufficient.","section":"Table 2"},{"comment":"The user fusion feature h_u,f is defined in Eq. (15) by aggregating h_i,m rather than h_i,f. As written, the user-side fusion representation is not actually formed, even though the later modality-aware preference module in Eqs. (18)-(21) consumes H_f. This appears to be a typo, but it needs correction and the implementation should be checked to confirm that fusion features are propagated to users.","section":"§4.2.1, Eq. (15)"}],"minor_comments":[{"comment":"The notation ¥S^m_a and ¥S^{m'}_b in Eq. (11) should presumably be ¥S^m_{a,b} and ¥S^{m'}_{a,b}; as written, the max appears to be taken over row or column vectors rather than scalar edge weights.","section":"§4.2.1, Eq. (11)"},{"comment":"The statement that the FFT reduces complexity to 'a logarithmic scale' is imprecise; the FFT complexity is O(n log n), not O(log n). Please rephrase.","section":"§4.1, Eq. (3)"},{"comment":"The abbreviation 'STOA' should be 'SOTA' for state-of-the-art.","section":"§5.1.2"}],"recommendation":"major_revision","confidential_remarks":"I have not independently verified the public GitHub repository. The main risk is Eq. (5): if the released code uses the filtered spectra in the fusion path, then the equation is a typo and the authors can correct it, rerun the affected ablation, and resubmit. If the code indeed uses unfiltered spectra, the central denoising mechanism is not implemented as described and the paper's contribution needs substantial reworking. The modest but consistent gains in Table 2 are plausible, but the missing variance reporting is a barrier to evaluating the central claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi [Colleague],\n\nThis one is worth a look, but with a caveat. The novel piece is the idea of moving modality fusion into the frequency domain: FFT on each modality's features, a learned complex filter, pointwise product for fusion, then inverse FFT. Combining that with graph propagation and a modality-aware preference module is genuinely new in the multimodal recommendation literature. The experiments are solid in scope — three datasets, five seeds, several baselines, and a public code URL — and the gains, while modest, are consistent.\n\nThe soft spot is the mechanism. Eq. (4) applies a trainable filter to each modality's spectrum to suppress noise. But Eq. (5), as written, defines the fusion spectrum as the pointwise product of the unfiltered spectra, with only a post-hoc fusion filter applied. So the per-modality denoising is bypassed exactly where the paper says it matters: fusion. This could simply be a typo — the intended product is probably over bH_i,m — but as published, the equations don't support the denoising story. The authors need to correct this and confirm which variant was actually evaluated in the experiments.\n\nThere are two smaller issues. Error bars are absent; the claim of p<0.01 in a paired t-test is made without reporting variances or effect sizes. And the denoising evidence is t-SNE plots, which are qualitative. No quantitative measure of noise reduction is given. The filter also has no explicit denoising objective, so improved accuracy could come from added capacity rather than genuine denoising.\n\nWho is this for? Researchers working on multimodal recommendation who want to explore spectral fusion. It's not a field reorientation, but it's a reasonable new tool. The paper deserves peer review — the idea is novel and the architecture is implementable — but the referee needs to push for the equation fix, error bars, and either a quantitative denoising test or a more modest framing.\n\nMy recommendation: engage with it, send to review, but flag the Eq. (5) issue as a necessary revision.","headline":"Novel frequency-domain fusion for multimodal rec, but the main denoising claim is undercut by an equation inconsistency that must be fixed.","tokens_in":15700,"tokens_out":2570,"would_cite":false,"duration_ms":15750,"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":"SMORE fuses item text and image features in the frequency domain, where a trainable filter suppresses modality noise, and reports consistent gains over nine baselines on three Amazon datasets.","keywords":["multimodal recommendation","spectrum-based fusion","Fourier transform","modality noise suppression","graph convolutional network","collaborative filtering","modality-aware preference","contrastive learning"],"falsifier":"Compare SMORE's Recall@20 against a variant in which the learned complex filter is replaced by an identity or an equally sized spatial-domain filter: if accuracy barely changes, the frequency-domain denoising mechanism is not the explanation for the gains. A more direct test is to inject known noise into modality features, pass them through the trained filter, and measure whether noise energy decays more than signal energy in the output spectra.","tokens_in":14712,"feed_emoji":"📡","tokens_out":6012,"duration_ms":54302,"temperature":0.7,"pith_summary":"The paper proposes SMORE, a multimodal recommender that fuses item text and image features in the frequency domain rather than concatenating or summing them. It claims that a trainable complex filter applied to Fourier-transformed features suppresses modality-specific noise, preventing the cross-modality noise amplification that direct fusion causes. SMORE also learns item-item graphs from both uni-modal and fused features, and balances uni-modal and fusion preferences with behavioral gating. On three Amazon datasets (Baby, Sports, Clothing), the paper reports that SMORE outperforms both general and multimodal baselines on Recall@K and NDCG@K. If the claim holds, spectrum-based fusion is a viable denoising mechanism for multimodal recommendation.","feed_headline":"Frequency-domain fusion lifts multimodal recommendation accuracy","feed_subtitle":"SMORE's spectral filter suppresses modality noise, beating general and multimodal baselines on Baby, Sports, and Clothing.","key_machinery":"The load-bearing device is the adaptive spectral filter: after an MLP projects each modality into a shared space, the fast Fourier transform converts the features into complex spectra, and a modality-specific trainable complex filter $\\delta_m$ attenuates noise via pointwise multiplication in Eq. (4). The filtered uni-modal spectra are then fused pointwise in Eq. (5), an operation equivalent to circular convolution in the spatial domain, capturing cross-modality correlations at logarithmic cost. The inverse FFT returns the denoised uni-modal and fused features to the original space for downstream graph learning and preference modeling.","core_discovery":"SMORE's central claim is that fusing modalities after projecting them into the frequency domain, filtering each spectrum with a trainable complex weight, and combining the filtered spectra by pointwise product recovers cross-modal correlations while attenuating modality-specific noise. The pointwise product in the frequency domain is equivalent to circular convolution in the spatial domain, so the fusion captures interactions between text and image features. On top of this, SMORE constructs separate item-item graphs for each modality and a fusion graph by max-pooling edge weights across modalities, then propagates denoised features through light graph convolution. A modality-aware preference module gates the uni-modal and fused features with behavioral signals and applies an InfoNCE contrastive loss to align behavioral and modality views. The paper reports consistent improvements over nine baselines on three real-world datasets, with ablations showing each component contributes.","pith_inferences":["The paper's evidence for denoising is indirect (accuracy and t-SNE); a direct measurement of spectra before and after filtering would test whether the filter is actually removing noise energy.","The spectral filtering idea could transfer to other multimodal tasks beyond recommendation, such as retrieval or classification, wherever pretrained encoders inject per-modality noise.","Since the fusion graph is built by max-pooling modality edge weights, a learned or attention-weighted fusion graph might change the balance between uni-modal and fusion signals; the paper does not compare these variants."],"forward_implications":["Direct fusion methods (concatenation, element-wise sum, attention) risk amplifying modality-specific noise; SMORE presents a form of fusion designed to suppress it.","The same spectral fusion module can extend beyond two modalities, since the pointwise product and per-modality filters apply to any number of feature streams.","Because FFT and pointwise aggregation cost logarithmic time, the fusion step avoids the quadratic cost of co-attention-based fusion.","Explicitly modeling uni-modal and fusion preferences, then balancing them with behavioral gating, is presented as a more realistic model of user choice than single fused representations."],"supporting_citations":[{"why":"LATTICE, the source of the item-item latent structure construction that SMORE extends and a primary multimodal baseline.","marker":"[42]"},{"why":"LightGCN, whose light graph convolution rule SMORE adopts for propagation.","marker":"[12]"},{"why":"VBPR, the early visual multimodal recommender used as baseline and dataset source.","marker":"[11]"},{"why":"MGCN, a baseline reducing unimodal noise via behavioral integration that SMORE compares against.","marker":"[41]"},{"why":"FREEDOM, a baseline that denoises the interaction graph and is outperformed by SMORE.","marker":"[50]"},{"why":"BM3, a self-supervised multimodal baseline in the comparison.","marker":"[52]"},{"why":"Provides the FFT machinery used to project features into the frequency domain.","marker":"[13]"},{"why":"InfoNCE contrastive loss that aligns behavioral and modality views in SMORE.","marker":"[24]"}],"fun_headline_variants":["Spectral fusion graph network cuts noise in multimodal recommendations","SMORE: Frequency-domain fusion suppresses noise, boosts recommendation","Graph-based spectral fusion improves multimodal recommendation accuracy","Fusing modalities in frequency domain reduces noise, lifts performance","Frequency-domain graph network filters noise for sharper recommendations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the trainable complex filter genuinely separates useful semantic content from modality-specific noise in the frequency domain, so that filtered spectra fused pointwise yield cleaner features.","fun_headline_variants_meta":{"raw":{"variants":["Spectral fusion graph network cuts noise in multimodal recommendations","SMORE: Frequency-domain fusion suppresses noise, boosts recommendation","Graph-based spectral fusion improves multimodal recommendation accuracy","Fusing modalities in frequency domain reduces noise, lifts performance","Frequency-domain graph network filters noise for sharper recommendations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000167,"raw_usage":{"total_tokens":1285,"prompt_tokens":998,"completion_tokens":287,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":614,"completion_tokens_details":{"reasoning_tokens":212}},"tokens_in":614,"tokens_out":287,"duration_ms":3257,"temperature":1.0,"reasoning_tokens":212,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:43:45.822109+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare SMORE's Recall@20 against a variant in which the learned complex filter is replaced by an identity or an equally sized spatial-domain filter: if accuracy barely changes, the frequency-domain denoising mechanism is not the explanation for the gains. A more direct test is to inject known noise into modality features, pass them through the trained filter, and measure whether noise energy decays more than signal energy in the output spectra.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"LightGCN, whose light graph convolution rule SMORE adopts for propagation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"VBPR, the early visual multimodal recommender used as baseline and dataset source."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MGCN, a baseline reducing unimodal noise via behavioral integration that SMORE compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"FREEDOM, a baseline that denoises the interaction graph and is outperformed by SMORE."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the FFT machinery used to project features into the frequency domain."}],"review_version":1}