{"id":"ea7bbfa3-0b26-4019-9a50-7e59b7dc73f5","arxiv_id":"2411.19822","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"SDR-GNN is a graph neural network that reconstructs missing multimodal features and labels utterance emotions, with reported gains over prior methods that are inconsistent across datasets.","lead":"A new graph network, SDR-GNN, aims to recognize emotions in conversations when audio, visual, or text data is missing. It combines hypergraph and gated aggregation to reconstruct missing modalities, reporting mixed but often improved accuracy on three benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Internal inconsistency in reported results: Table 2 and Table 5 give different WAF1 for the same SDR-GNN configuration, so the SOTA claim is not currently supported.","rationale":"The paper's central claim is empirical: SDR-GNN outperforms state-of-the-art methods on incomplete multimodal conversational emotion recognition. The strongest evidence for this claim is the set of tables reporting WAF1 and ACC across missing rates. Table 2 and Table 5 present different numbers for the same condition (IEMOCAP four-class, missing rate 0.6), and baselines also shift between tables. This is not a minor formatting issue; it means the paper's own evidence cannot be reconciled. The reader's weakest_assumption focused on the modeling constraint that at least one modality is always available, which is a real limitation, but the table inconsistency is more immediately load-bearing because it calls into question every quantitative result. My read therefore supports the reader's REJECT verdict without changing it; the verdict should remain as it was.","tokens_in":22155,"tokens_out":2663,"duration_ms":22518,"concrete_test":"Run the released code (github.com/fufangze/SDR-GNN) on IEMOCAP four-class at missing rate 0.6, using the Section 4.2 hyperparameters, and compare the resulting WAF1 with Table 2 (75.03) and Table 5 (81.13). If the reproduced value does not match both within a small tolerance, the reported results are not reproducible and the SOTA claim cannot be accepted. Also re-run missing rate 0.0 to check whether reconstruction at 60% loss can plausibly beat complete data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.4's Table 5 reports SDR-GNN on IEMOCAP four-class at missing rate 0.6 with WAF1=81.13, while Table 2 reports Ours=75.03 for the same dataset and missing rate. Baselines differ as well: GCNet is 78.87 in Table 5 but 71.88 in Table 2, and CPM-Net is 68.68 in Table 5 but 47.38 in Table 2. The two tables describe the same experimental setting (IEMOCAP four-class, missing rate 0.6), so at least one set of numbers is wrong, or the configurations differ in a way the paper does not disclose. Note Table 5's 81.13 is even higher than Table 2's value at missing rate 0.0 (79.58), which would imply that losing 60% of modalities improves accuracy over complete data. Since the abstract's central claim is that SDR-GNN 'outperforms current state-of-the-art methods,' and Tables 2-3 are the only direct evidence, this internal inconsistency undermines the core empirical claim. It is not a question of whether the mechanism name is apt; the reported measurements themselves are not self-consistent.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SDR-GNN, a graph neural network for incomplete multimodal emotion recognition in conversations. SDR-GNN builds speaker and context interaction graphs over utterances, adds a hypergraph and a learned gating mechanism described as frequency-aware, reconstructs missing modality features, and classifies emotions. It is evaluated on IEMOCAP, CMU-MOSI, and CMU-MOSEI at missing rates from 0.0 to 0.7, reporting weighted F1 and accuracy, and the abstract claims that the method \"outperforms current state-of-the-art methods.\"","tokens_in":22476,"tokens_out":5406,"duration_ms":46003,"significance":"If the reported results were reliable, SDR-GNN would be a meaningful contribution to incomplete multimodal learning in conversational emotion recognition. The paper has notable strengths: the authors provide public code, a detailed ablation study (Table 4), explicit parameter tuning (Section 5.7), a reconstruction-error comparison (Figure 5), and a case study (Figure 8). However, the central empirical claim is not currently supported: the paper's own tables contain a direct numerical contradiction, and several baselines actually beat the proposed method on CMU-MOSI. The spectral/frequency interpretation of the core mechanism is also not demonstrated. These issues are load-bearing for the claimed contribution.","major_comments":[{"comment":"There is a direct internal inconsistency in the reported results. For IEMOCAP four-class at missing rate 0.6, Table 2 gives SDR-GNN WAF1=75.03, GCNet=71.88, and CPM-Net=47.38, while Table 5 gives SDR-GNN=81.13, GCNet=78.87, and CPM-Net=68.68 for the same dataset, same missing rate, and same metric. At least one set of numbers is wrong, or the experimental configurations differ in a way the paper does not disclose. Since the abstract's SOTA claim rests on Tables 2 and 3, this contradiction undermines the central empirical result and must be resolved before the paper can be evaluated.","section":"Table 2 vs Table 5"},{"comment":"In the paper's own comparison, DiCMoR outperforms SDR-GNN on CMU-MOSI at missing rates 0.2 (82.0/82.1 vs. 81.9/81.9), 0.5 (76.4/76.7 vs. 76.1/76.2), and 0.6 (73.0/73.3 vs. 72.2/72.2). This contradicts the abstract's claim that SDR-GNN outperforms current state-of-the-art methods. The authors should either qualify the claim (e.g., \"outperforms on IEMOCAP and CMU-MOSEI\") or provide statistical evidence, such as significance tests over repeated runs, that supports an overall superiority claim.","section":"Table 3, CMU-MOSI"},{"comment":"The baseline numbers are largely borrowed from previous papers: the footnote to Table 2 states \"results come from [22]\" and Table 3 states \"results come from [60].\" These numbers were obtained under different experimental protocols, including different masking implementations, random seeds, and possibly different feature preprocessing or data splits. The paper does not rerun any baseline under its own masking protocol, and it reports no standard deviation or confidence interval despite stating in Section 4.2 that performance is averaged over ten trials. Without a shared protocol and variance reporting, the comparative claims in Tables 2 and 3 are not quantitatively meaningful.","section":"Tables 2-3, baselines"},{"comment":"The paper's central novelty is described as \"Spectral Domain Reconstruction\" and \"multi-frequency aggregation,\" but Equations (8) and (9) implement a learned gating mechanism: a tanh-scaled linear function of concatenated node features. No graph Laplacian, no spectral decomposition, and no frequency-domain filtering appears in the model. The statement that \"These signals are derived through the spectral decomposition of the graph Laplacian\" (Section 3.2) is not matched by any mathematical or experimental analysis of the learned filters. The authors should either provide evidence that the gating mechanism actually separates high- and low-frequency graph signal components, or reframe the contribution as a gated aggregation method without the spectral claim.","section":"Section 3.2, Eqs. (8)-(9)"},{"comment":"The evaluation is restricted to a specific missingness model: randomly masking individual utterance modalities while ensuring at least one modality remains per utterance. Section 3.1 states this assumption explicitly, and Section 4.2 confirms that rho ranges only to 0.7, the approximate maximum under the at-least-one-modality constraint. Real conversations can contain utterances with all modalities missing (e.g., sensor failure) or structured missing patterns (e.g., an entire speaker's audio channel absent). The paper does not discuss or test these cases, so the claimed robustness of the reconstruction layer (Eq. 11) is bounded by this assumption. The authors should state this limitation in the abstract and conclusion and, ideally, evaluate a setting with all-modality-missing utterances.","section":"Section 3.1 and Section 4.2"}],"minor_comments":[{"comment":"The text says \"in Fig. 3, we compare the performance\" but Figure 3 shows confusion matrices; the correct cross-reference appears to be Figure 4, which shows the lower-bound comparison.","section":"Section 5.5"},{"comment":"The text refers to \"Fig. 6\" for the reconstruction-loss weight e and \"Fig. 7\" for the number of layers l, but the figure captions are reversed: Figure 6 is captioned \"Parameter tuning with various number of hypergraph\" and Figure 7 is captioned \"various weight of reconstruction loss function.\" The cross-references must be fixed.","section":"Section 5.7"},{"comment":"There are typos in the ablation discussion: \"spearker\" should be \"speaker,\" and \"Codeceases\" appears to be a typo for \"decreases.\" Also, Section 5.2 defines SDR-GNNw/o Fre as replacing the frequency-aware layer with DialogueGCN, but the text later says \"we replace the frequency-aware graph learning with a graph convolution network from DialogueGCN,\" which is clear after the typo fix.","section":"Section 5.2"},{"comment":"The reconstruction loss computes the MSE between the reconstructed and original features, but when a modality is missing, the original is not observed in the test setting. The paper should clarify that this loss is computed only during training where the missingness is artificially simulated, since otherwise Eq. (16) would require access to the ground-truth missing features.","section":"Section 3.3, Eq. (16)"},{"comment":"The incidence matrix in Eq. (3) is denoted with a hat as \\hat{H} and described as a weighted incidence matrix, but Eq. (6) later uses both H and \\hat{H}; the relationship between the unweighted and weighted matrices should be defined explicitly to avoid ambiguity.","section":"Section 3.2, Eq. (3)"}],"recommendation":"major_revision","confidential_remarks":"The contradiction between Table 2 and Table 5 is serious and should be investigated carefully; it may indicate a reporting error, a configuration difference, or a deeper issue with the experimental pipeline. The editor may wish to request the raw per-run results and the exact experimental configuration for both tables. Additionally, many of the references are self-citations to recent arXiv preprints by the same group; while not a problem by itself, the authors should ensure that all related prior work is cited accurately and that the novelty claims are positioned against it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the central SOTA claim doesn't survive contact with the paper's own tables. IEMOCAP four-class at r=0.6 is 75.03 WAF1 in Table 2 and 81.13 in Table 5, and baselines shift too (GCNet 71.88 vs 78.87). On CMU-MOSI, DiCMoR beats SDR-GNN at r=0.2, 0.5, and 0.6. So the claim that SDR-GNN consistently outperforms current methods is not supported by the reported numbers.\n\nWhat's actually new: the architecture—R-GCN plus hypergraph with edge-dependent node weights, a gated frequency-style aggregation, and a reconstruction loss—applied to incomplete multimodal ERC. That combination isn't in the cited prior work. The ablations are reasonable and each component appears to earn its place. The code is public, which is good.\n\nThe soft spots, in proportion: the 'spectral domain' language overstates the math; Eqs. 8-9 are a learned gate, not spectral filtering. Baselines are borrowed with no shared masking protocol and no variance, despite a claim of ten trials. The assumption that at least one modality is present and masking is uniform random is explicit and reasonable for a first pass.\n\nWho should read it: anyone working on IMERC or graph-based reconstruction for multimodal conversation. It deserves peer review because the method is sensible and the datasets are standard, but the empirical reporting needs major revision before the results can be trusted. I wouldn't cite it in this state.","headline":"A plausible GNN combination for incomplete multimodal ERC, but the SOTA claim is undermined by an internal table conflict and by the paper's own MOSI numbers.","tokens_in":22998,"tokens_out":3922,"would_cite":false,"duration_ms":31665,"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":"SDR-GNN claims that a spectral-domain graph network, by keeping high-frequency signals alive during aggregation, can reconstruct missing modalities and sustain conversational emotion recognition performance up to 70% missing rates.","keywords":["Incomplete multimodal learning","Conversational emotion recognition","Multimodal fusion","Spectral domain reconstruction","Graph neural network","Hypergraph","High-frequency information","Modality imputation"],"falsifier":"Run SDR-GNN on IEMOCAP with a test set where one speaker's audio is entirely removed while text and video stay complete, matching the same overall missing proportion as a uniform-random condition, and compare WAF1. If performance falls to the level of a no-audio unimodal baseline, the claim that multi-frequency aggregation alone drives robustness fails; a second check is to train with some utterances having all three modalities masked and observe whether the reconstruction layer can produce any signal at all.","tokens_in":1685,"feed_emoji":"🎭","tokens_out":3300,"duration_ms":54272,"temperature":0.7,"pith_summary":"The paper tries to establish that a graph neural network operating in the spectral domain can recover missing audio, visual, and text features in conversation, so that emotion recognition stays accurate even when most modality data is absent. It argues that previous graph methods lose high-frequency signals—the sharp emotional shifts that distinguish anger from neutral, for example—because message passing smooths them away. SDR-GNN builds a speaker-and-context interaction graph, enriches it with a weighted hypergraph that captures multi-node relationships, and adds a self-gating mechanism that keeps both low- and high-frequency information during aggregation. It then reconstructs missing modalities with a linear layer and fuses everything through multi-head attention. Across IEMOCAP, CMU-MOSI, and CMU-MOSEI, the authors report consistent gains over comparison methods, with the largest advantages at high missing rates.","feed_headline":"Graph network rebuilds missing audio and video to read emotions","feed_subtitle":"Spectral aggregation keeps high-frequency cues alive, so emotion recognition holds up even when 70% of modalities are missing.","key_machinery":"The central mechanism is a graph built from a sliding context window over utterances, with two edge types—speaker identity edges and context-position edges—that feed relation-aware graph convolutions. On top of this sits a weighted hypergraph with edge-dependent node weights, which lets one hyperedge connect multiple utterances and capture higher-order dependencies. The frequency-aware module is a self-gating aggregation: for each edge, a gating score computed from the concatenated node pair determines whether low-frequency (similarity) or high-frequency (difference) signals dominate, using tanh to scale the gate between -1 and 1. This gating drives the reconstruction layer, which predicts missing modality features from the learned representations, and the multi-head attention layer that refines the reconstructed data.","core_discovery":"The paper's central claim is that SDR-GNN outperforms current state-of-the-art methods for incomplete multimodal emotion recognition in conversations, on all three benchmark datasets and across missing rates from 0.0 to 0.7. The reported gains are 0.77% to 8.6% in weighted average F1 over baselines, and the model degrades less than others as the missing rate rises: on IEMOCAP four-class, its WAF1 drops by only 5.17% from missing rate 0.0 to 0.7, while comparison methods drop between 6.98% and 37.70%. The paper attributes this robustness to two design choices: a weighted hypergraph that models higher-order utterance dependencies beyond binary edges, and a frequency-aware aggregation that preserves high-frequency information that ordinary graph convolutions over-smooth away. The reconstruction layer then uses the multi-frequency-enriched features to impute missing modalities, and a multi-head attention layer optimizes the recovered data before classification.","pith_inferences":["The uniform-random masking assumption likely overstates real-world robustness; correlated missing patterns, such as one speaker's audio being entirely absent, should be tested to see whether the reconstruction layer can still interpolate from a missing-stream context.","The self-gating mechanism might transfer to other conversational tasks where abrupt state changes matter, such as sarcasm detection or dialogue-act shift detection, because it explicitly preserves node-level differences that standard smoothing erases.","Since the reconstruction loss weight $e=0.5$ was found optimal, an adaptive weighting scheme that increases reconstruction pressure as the missing rate grows could extend the method beyond 0.7, although the at-least-one-modality constraint would still bind.","The paper's attribution of gains to high-frequency preservation could be tested directly by flipping the sign of the gating scores (making the model favor smoothing), which should sharply reduce performance at high missing rates if the mechanism is load-bearing."],"forward_implications":["At missing rate 0.7, SDR-GNN declines only 5.17% on IEMOCAP four-class, while baselines drop between 6.98% and 37.70%, so the model's robustness gap widens precisely where data is most incomplete.","On fully complete data (missing rate 0.0), SDR-GNN still improves WAF1 by 0.7% to 2.7% over baselines, meaning the multi-frequency architecture helps even without any missing modalities.","Reconstruction quality, measured by MSE between imputed and true features, stays lowest for SDR-GNN and grows more slowly as the missing rate increases, indicating that the frequency-aware features genuinely improve imputation.","Discarding incomplete utterances (the lower-bound setting) consistently hurts performance compared with using them, so the incomplete data itself carries signal that SDR-GNN exploits.","The case study shows SDR-GNN and MMIN preserve a high-frequency emotional transition (a shift from neutral to anger) that other context-based models miss, which supports the claim that high-frequency preservation matters for emotion recognition."],"supporting_citations":[{"why":"Supplies the GCNet baseline and the reported comparison numbers for many methods; it is the primary graph-completion competitor SDR-GNN must beat.","marker":"[22]"},{"why":"DiCMoR, the distribution-consistent modality-recovery baseline that provides the MOSI/MOSEI comparison table SDR-GNN is evaluated against.","marker":"[60]"},{"why":"Establishes that multivariate, multi-frequency graph information benefits emotion recognition in conversation, motivating the frequency-aware aggregation module.","marker":"[16]"},{"why":"Diagnoses the low-pass filtering and over-smoothing of graph convolutions, the problem the frequency-aware self-gating is designed to solve.","marker":"[12]"},{"why":"MCTN, the cyclic-translation reconstruction baseline used for comparison on CMU-MOSI and CMU-MOSEI.","marker":"[33]"},{"why":"MMIN, the missing-modality imagination baseline that handles six missing patterns and provides a strong comparison point on all three datasets.","marker":"[68]"}],"fun_headline_variants":["Spectral graph preserves high-frequency cues for emotion recognition","Missing modalities? Spectral graph keeps emotion recognition sharp","Graph network with spectral aggregation handles incomplete data","SDR-GNN: reconstructs missing modalities using high-frequency signals","High-frequency graph imputation boosts emotion recognition under gaps"],"cache_read_input_tokens":25088,"weakest_assumption_plain":"Every utterance keeps at least one of its three modalities, and missing modalities are chosen randomly with uniform probability; if an utterance loses all modalities, or missingness is patterned (like one speaker's audio always gone), the reconstruction layer has nothing to work from and the claimed robustness is not guaranteed.","fun_headline_variants_meta":{"raw":{"variants":["Spectral graph preserves high-frequency cues for emotion recognition","Missing modalities? Spectral graph keeps emotion recognition sharp","Graph network with spectral aggregation handles incomplete data","SDR-GNN: reconstructs missing modalities using high-frequency signals","High-frequency graph imputation boosts emotion recognition under gaps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000222,"raw_usage":{"total_tokens":1486,"prompt_tokens":1008,"completion_tokens":478,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":624,"completion_tokens_details":{"reasoning_tokens":402}},"tokens_in":624,"tokens_out":478,"duration_ms":4932,"temperature":1.0,"reasoning_tokens":402,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:46:59.626429+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SDR-GNN on IEMOCAP with a test set where one speaker's audio is entirely removed while text and video stay complete, matching the same overall missing proportion as a uniform-random condition, and compare WAF1. If performance falls to the level of a no-audio unimodal baseline, the claim that multi-frequency aggregation alone drives robustness fails; a second check is to train with some utterances having all three modalities masked and observe whether the reconstruction layer can produce any signal at all.","supporting_citations":[{"cited_title":"Gcnet:Graphcom- pletion network for incomplete multimodal learning in conversation","cited_arxiv_id":null,"evidence_quote":"Supplies the GCNet baseline and the reported comparison numbers for many methods; it is the primary graph-completion competitor SDR-GNN must beat."},{"cited_title":"Distribution-consistent modal recoveringforincompletemultimodallearning,in:Proceedingsofthe IEEE/CVFInternationalConferenceonComputerVision(ICCV),pp","cited_arxiv_id":null,"evidence_quote":"DiCMoR, the distribution-consistent modality-recovery baseline that provides the MOSI/MOSEI comparison table SDR-GNN is evaluated against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes that multivariate, multi-frequency graph information benefits emotion recognition in conversation, motivating the frequency-aware aggregation module."},{"cited_title":"Beyond low-frequency information in graph convolutional networks, in: Proceedings of the AAAI conference on artificial intelligence, pp","cited_arxiv_id":null,"evidence_quote":"Diagnoses the low-pass filtering and over-smoothing of graph convolutions, the problem the frequency-aware self-gating is designed to solve."},{"cited_title":"Found in translation: Learning robust joint representations by cyclic translationsbetweenmodalities,in:ProceedingsoftheAAAIconfer- ence on artificial intelligence, pp","cited_arxiv_id":null,"evidence_quote":"MCTN, the cyclic-translation reconstruction baseline used for comparison on CMU-MOSI and CMU-MOSEI."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MMIN, the missing-modality imagination baseline that handles six missing patterns and provides a strong comparison point on all three datasets."}],"review_version":1}