{"id":"b1c9563f-584d-4fe9-9b48-39870f74be08","arxiv_id":"2501.14246","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A three-expert progressive attention graph neural network improves EEG emotion classification accuracy on SEED, SEED-IV, and MPED benchmarks.","lead":"This paper introduces a neural network with three attention stages that reads brain-wave (EEG) signals and learns which electrode locations matter for recognizing emotions. It reports accuracy gains over earlier models on three standard emotion datasets, with the largest gains on the SEED-IV four-class task.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never states whether the class index l in Eq. (5) is the true label or the model's predicted label when computing attention maps at inference; if true labels are used on test data, the reported accuracies in Table I are invalid, and if predicted labels are used, the train/test procedure…","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern that I would raise: the undefined label l in Eq. (5) determines whether the progressive attention maps used at inference leak ground-truth labels. This is the most direct threat to the paper's central empirical claim because it concerns the validity of the test-time evaluation itself, not just the interpretation of which component caused the gain. The paper's ablations (Tables II-IV) do support the existence of a progressive-attention benefit, and the diversity loss and dynamic fusion components are plausible, but none of that matters if the test-time attention maps are computed with access to the true emotion labels. I agree with the reader's CONDITIONAL verdict: the concern is severe enough to block acceptance without clarification, but it is addressable by stating the inference-time label rule and, ideally, by releasing code or running a predicted-label ablation. I considered whether the parameter-count confound in the 2E-versus-3E comparison should be the primary concern, but that affects only the causal attribution to progressive attention, whereas the Eq. (5) label ambiguity affects the validity of every reported accuracy. Therefore no verdict change is needed relative to the reader's conditional assessment.","tokens_in":14728,"tokens_out":5613,"duration_ms":54988,"concrete_test":"Obtain the authors' code (or ask them to release it) and inspect the inference path for Eqs. (5)-(7): determine exactly what value of l is used to compute α_l^d for test samples. Then, if the code is not available, re-implement APAGNN from the paper and run the SEED and SEED-IV test sets under two conditions: (A) l is set to the ground-truth label, and (B) l is set to the argmax of the expert's predicted class probability. Compare the resulting accuracies with Table I. If conditions (A) and (B) differ by more than about 0.5% absolute accuracy, the missing protocol statement is empirically consequential. If the code already uses condition (B), the practical concern is resolved, but the manuscript must still state that training uses true labels and inference uses predicted labels, and the experiments should be re-run under that stated protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The core empirical claim is that APAGNN reaches 96.38% on SEED, 86.64% on SEED-IV, and 41.58% on MPED. The mechanism behind this claim is progressive attention: expert 1 produces an attention map Φ1 that prunes nodes for expert 2, and expert 2 produces Φ2 that prunes nodes for expert 3 (Section III-A). The attention maps are computed from Eq. (5), α_l^d = (1/C) Σ_c ∂S_i^l / ∂H_i, and Eq. (6), I_i = ReLU(Σ_d α_l^d H_i). This requires choosing a 'target emotion class index l'. The paper never specifies what l is at inference time. If l is the ground-truth label, then the test-time graph seen by experts 2 and 3 is constructed using the answer key, which is a direct label leak that would invalidate every number in Table I. If l is the predicted label (e.g., argmax of S_i), the method is legitimate but the paper still must state that training and inference use different label sources, since the training loss in Eq. (4) uses the true label. The manuscript also provides no code, so the ambiguity cannot be resolved from the text. This is not a stylistic omission: it is the single most load-bearing assumption separating a valid evaluation from an invalid one. A secondary concern is that the 3E-versus-2E ablation in Table II does not equalize parameter count, so the causal attribution to 'progressive attention' is weaker than the accuracy claim; however, the label-protocol issue is more fundamental because it affects the validity of the headline accuracies themselves.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes APAGNN, a multi-expert graph neural network for EEG emotion recognition. Three experts process the EEG graph at progressively finer granularity: the first captures global topology, the second region-level patterns, and the third electrode-level features; the first two experts generate gradient-based attention maps (Eqs. 5-6) that prune channels for the next expert. A Jensen-Shannon diversity loss encourages the experts to focus on different patterns, and a learned weight generator fuses their representations. Experiments on SEED, SEED-IV, and MPED report accuracies of 96.38%, 86.64%, and 41.58%, respectively, together with ablations on the number of experts, the diversity loss, static versus adaptive attention, confusion matrices, attention visualizations, and t-SNE plots.","tokens_in":15104,"tokens_out":3735,"duration_ms":36635,"significance":"If the reported results are trustworthy, the paper makes a useful empirical contribution: adaptive, subject-specific channel selection through progressive attention is a plausible and interpretable way to improve EEG emotion classification, and the comparison against static channel sets (Table IV) speaks directly to that design choice. The paper also includes several complementary analyses (confusion matrices, attention visualizations, t-SNE) that go beyond a single accuracy number. However, the central empirical claim is currently gated by an unresolved test-time protocol issue involving the class index in Eq. (5), and the headline comparison is partially undercut by missing baseline entries in Table I and by non-matched ablations. The paper ships no code, so the protocol ambiguity cannot be resolved by inspection.","major_comments":[{"comment":"The attention maps used to prune nodes for experts 2 and 3 require a target emotion class index l in Eq. (5), but the manuscript never states what l is at inference time. If l is the ground-truth label during testing, then the graph seen by the later experts is constructed using the answer key, which is a direct label leak and would invalidate every accuracy in Table I. If l is the model's predicted label (for example, argmax of S_i), the procedure is legitimate but must be stated, and the fact that training (Eq. (4)) uses true labels while inference would use predicted labels must be discussed and tested. This is the most load-bearing unresolved detail in the paper; please specify the protocol and, if necessary, rerun the experiments under the stated protocol.","section":"Section III-A, Eqs. (5)-(6)"},{"comment":"The text claims APAGNN 'achieves superior performance compared to the existing methods on all three datasets,' but Table I has missing entries: SEED-IV is blank for both RGNN and IAG, and MPED is blank for RGNN. The subsequent sentence comparing with 'the previous best method, IAG' is also inconsistent with the table, since V-IAG reports higher accuracy than IAG on both SEED (95.64 vs 95.44) and MPED (40.40 vs 40.38). Complete the baseline table or revise the claim.","section":"Table I and Section IV-C"},{"comment":"The ablation APAGNN-2E versus APAGNN-3E does not equalize parameter count, computational cost, or training objective complexity; the two architectures differ in the number of expert branches and in the number of attention-transfer stages. The observed 2.54-3.05 percentage point gains are therefore not a controlled test of the 'progressive attention' mechanism, and the causal attribution to progressive attention is weaker than the text suggests. A parameter-matched control, or a discussion of this confound, is needed.","section":"Table II and Section IV-C"},{"comment":"No significance tests or repeated-run statistics are reported. For example, the SEED gap between APAGNN and V-IAG is 0.74 percentage points (96.38±4.19 versus 95.64±5.08), which is much smaller than the reported standard deviations, and it is unclear whether the per-subject or per-trial comparisons are paired. Without significance testing, the claim of 'superior performance' is not statistically supported even setting aside the protocol issue.","section":"Tables I-III and Section IV-C"}],"minor_comments":[{"comment":"There is a typo: 'we conducte extensive experiments' should be 'we conducted extensive experiments.'","section":"Section IV-A"},{"comment":"The caption states that 'only the first two are implemented in the APAGNN,' which conflicts with Section III-A and Fig. 2, where each expert is described as performing attention-map generation. Please clarify whether the third expert generates an attention map or not.","section":"Fig. 4 caption"},{"comment":"The text says the diversity loss is important 'across datasets,' but Table III reports results only for SEED and SEED-IV; MPED is omitted. Either add the MPED result or qualify the claim.","section":"Section IV-C and Table III"},{"comment":"Several placeholders remain in the table of contents, e.g., 'Y .' in references has an extra space; please run a copyediting pass.","section":"Section IV-B"},{"comment":"No code or data availability statement is provided. Given the ambiguity in Eq. (5), releasing code (or at least a precise pseudocode for inference) would substantially strengthen the paper.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The label-protocol ambiguity in Eqs. (5)-(6) is the key gate for this paper. If the authors cannot confirm that predicted labels are used at inference, the results cannot be considered valid. I would ask for an explicit statement of the inference protocol and, ideally, a rerun of the affected experiments or a release of code before considering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: this is a coherent architecture paper with a real evaluation ambiguity that blocks trusting the headline numbers. Eq. (5) computes a Grad-CAM-style attention map for a 'target emotion class index l', and the paper never states where l comes from at inference. If it is the ground-truth label, the reported accuracies in Table I are invalid because the test-time graph for experts 2 and 3 is built using the answer key. If it is the model's predicted label, the procedure is legitimate but training and inference differ (Eq. (4) uses the true label), and that mismatch needs to be stated and its sensitivity tested. No code is provided, so the text alone cannot resolve this. This is the load-bearing issue.\n\nCredit where it's due: the progressive architecture — global expert, then region expert using the first attention map to prune the graph, then channel expert using the second — is a coherent and fairly new combination for EEG emotion recognition. The JS-divergence diversity loss between the first two experts is a sensible addition. The ablations in Tables II and III support the claims that more experts and the diversity loss help, and dynamic attention beats fixed channel sets (Table IV). The attention visualizations align with frontal and temporal regions, consistent with the emotion literature.\n\nSoft spots beyond the label ambiguity: Table I omits RGNN and IAG values for SEED-IV, so 'superior' is asserted against an incomplete baseline list. There are no significance tests or repeated-run confidence intervals, and the 3E-vs-2E ablation does not equalize parameter count, so the accuracy gain cannot be cleanly attributed to progressive attention alone. These are addressable issues, not fatal ones.\n\nWho this is for: researchers in affective computing and EEG-based GNNs. The method is not a breakthrough but is a solid experimental combination that could be useful if the evaluation is tightened. It deserves a serious referee: the central flaw is an omission that can be corrected. I would ask the authors to clarify the inference protocol, release code, fill in the baseline table, and run significance tests. If the numbers hold up, this becomes a decent contribution.","headline":"Plausible progressive-attention GNN for EEG emotion recognition, but an unspecified label source in the attention computation leaves the headline accuracies unverifiable.","tokens_in":15677,"tokens_out":2870,"would_cite":false,"duration_ms":24982,"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":"Three progressively focused graph experts set new accuracy marks for EEG emotion recognition on SEED, SEED-IV, and MPED.","keywords":["EEG emotion recognition","graph neural network","progressive attention","Grad-CAM","multi-expert learning","differential entropy","Jensen-Shannon divergence","subject-dependent classification"],"falsifier":"Re-run the SEED and SEED-IV experiments while logging which label is used to compute Eq. (5) at inference. If the true label is used on test samples, the 96.38% and 86.64% accuracies are not a valid measure of how the model would perform on unlabeled data; if the predicted label is used and the accuracies hold, the protocol concern is resolved.","tokens_in":14531,"feed_emoji":"🧠","tokens_out":4693,"duration_ms":39030,"temperature":0.7,"pith_summary":"The paper proposes APAGNN, a graph neural network with three expert modules that analyze EEG brain activity from global patterns down to individual electrodes. Each expert learns to classify emotions and generates an attention map that prunes less relevant EEG channels before the next expert sees the data. The authors claim this progressive attention design outperforms existing methods on three benchmark datasets, achieving 96.38% accuracy on SEED, 86.64% on SEED-IV, and 41.58% on MPED, with the lowest standard deviation across all three. A sympathetic reader would care because the attention maps offer a subject-specific view of which brain regions drive emotion classification.","feed_headline":"Three expert GNNs push EEG emotion accuracy to 96.38%","feed_subtitle":"Progressive attention prunes channels expert by expert, beating prior methods on SEED, SEED-IV, and MPED.","key_machinery":"The progressive attention mechanism is the load-bearing component. Expert $i$ computes a channel-importance map $I_i$ by combining the feature representation $H_i$ with pooled gradients of the predicted probability for a target class, then normalizes and thresholds it at $\\eta$ to prune weakly relevant electrode nodes and edges before the next expert. This lets later experts operate on a subgraph of the original electrode topology. The diversity-preserving loss maximizes the Jensen-Shannon divergence between the first two experts' normalized attention maps, and the dynamic fusion weight generator learns coefficients $\\xi_1, \\xi_2, \\xi_3$ to combine the three experts' representations into the final prediction.","core_discovery":"APAGNN's central claim is that progressive attention, not the graph convolution backbone alone, drives the performance gain. The model chains three experts: the first reads the full electrode graph, the second focuses on emotion-relevant regions selected by the first expert's attention map, and the third concentrates on the most informative electrodes. Each expert is a Chebyshev graph convolutional network on differential-entropy features, and the attention maps are computed with a Grad-CAM-style gradient of the predicted class score with respect to the feature map. A Jensen-Shannon divergence term forces the first two experts to attend to different channels, and a dynamic weight generator fuses the three experts' representations. On the subject-dependent splits of SEED, SEED-IV, and MPED, the model reports accuracies of 96.38%, 86.64%, and 41.58%, surpassing the previous best methods and achieving the lowest standard deviations.","pith_inferences":["The paper never states whether the attention map at test time uses the true emotion label or the model's own predicted label; if the true label is used, the reported accuracies are not a valid estimate of deployment performance. This is an inference from the missing protocol detail, not a paper claim.","The subject-dependent protocol, where training and test trials come from the same subjects, means the method's advantage over subject-independent approaches remains untested; one could extend the comparison to leave-one-subject-out evaluation.","The fixed pruning threshold $\\eta=0.5$ is a hyperparameter that could be learned or annealed, which might further improve the progressive refinement.","The dynamic weight generator effectively learns a soft mixture-of-experts, yet the paper does not analyze how the fusion weights vary across subjects or emotions."],"forward_implications":["If the central claim holds, progressively refined attention maps are a viable way to handle inter-subject variability in emotion-related brain topology.","The three-expert setup outperforms the two-expert variant by roughly 2.5 to 3 percentage points, suggesting additional refinement stages can keep adding accuracy.","Dynamic attention, which learns which channels matter per sample, beats static channel sets from prior neuroscience work by about 3 to 4 points on SEED and SEED-IV.","The attention maps highlight prefrontal and temporal regions consistent with emotion neuroscience, giving the model an interpretable output."],"supporting_citations":[{"why":"Grad-CAM supplies the gradient-based attention mechanism used to compute each expert's channel importance map.","marker":"[47]"},{"why":"Chebyshev spectral filtering provides the graph convolution operator that forms each expert's feature extractor.","marker":"[41]"},{"why":"DGCNN is a dynamic graph baseline that APAGNN must beat and whose approach it extends.","marker":"[39]"},{"why":"RGNN is a regularized GNN baseline and also the source of one predefined static channel set used in the attention comparisons.","marker":"[13]"},{"why":"IAG is the previous best method on SEED and MPED, and the paper's accuracy gains are measured against it.","marker":"[40]"},{"why":"Zheng and Lu provide the SEED dataset protocol and the predefined critical channel set used in the static-attention variant SPAGNN-v1.","marker":"[49]"},{"why":"Song et al. supply the MPED dataset and its train/test protocol used for the seven-class emotion recognition experiments.","marker":"[51]"}],"fun_headline_variants":["Progressive attention, not graph depth, boosts EEG emotion","Three-expert GNN prunes channels to lift emotion recognition","Grad-CAM attention steers progressive EEG emotion GNN","Hierarchical experts capture brain regions for emotion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported accuracies depend on an unspecified detail in the attention computation: whether the class used to generate each test-time attention map is the true emotion label or the model's own prediction, and the paper never states which one is used.","fun_headline_variants_meta":{"raw":{"variants":["Progressive attention, not graph depth, boosts EEG emotion","Three-expert GNN prunes channels to lift emotion recognition","Grad-CAM attention steers progressive EEG emotion GNN","Hierarchical experts capture brain regions for emotion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000199,"raw_usage":{"total_tokens":1334,"prompt_tokens":871,"completion_tokens":463,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":487,"completion_tokens_details":{"reasoning_tokens":397}},"tokens_in":487,"tokens_out":463,"duration_ms":4629,"temperature":1.0,"reasoning_tokens":397,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:14:48.990934+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the SEED and SEED-IV experiments while logging which label is used to compute Eq. (5) at inference. If the true label is used on test samples, the 96.38% and 86.64% accuracies are not a valid measure of how the model would perform on unlabeled data; if the predicted label is used and the accuracies hold, the protocol concern is resolved.","supporting_citations":[{"cited_title":"Eeg emotion recognition using dynamical graph convolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"DGCNN is a dynamic graph baseline that APAGNN must beat and whose approach it extends."},{"cited_title":"Eeg-based emotion recognition using regularized graph neural networks,","cited_arxiv_id":null,"evidence_quote":"RGNN is a regularized GNN baseline and also the source of one predefined static channel set used in the attention comparisons."},{"cited_title":"Instance-adaptive graph for eeg emotion recognition,","cited_arxiv_id":null,"evidence_quote":"IAG is the previous best method on SEED and MPED, and the paper's accuracy gains are measured against it."},{"cited_title":"Mped: A multi-modal physiological emotion database for discrete emotion recognition,","cited_arxiv_id":null,"evidence_quote":"Song et al. supply the MPED dataset and its train/test protocol used for the seven-class emotion recognition experiments."}],"review_version":1}