{"id":"fac1e944-c823-4eaa-b734-a9bf010a56d4","arxiv_id":"2412.19200","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DSAML uses dual-scale attention and meta-learning over per-annotator tasks to predict dynamic music emotion for both the average listener and for individual listeners from a single personalized song annotation.","lead":"This paper introduces a personalized approach to dynamic music emotion recognition, predicting how a specific listener would rate the emotion of a song over time. The method combines a dual-scale attention model with meta-learning so that one annotated song from a new user is enough to adapt predictions to that user's perception.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The one-shot PDMER claim assumes per-annotator labels are stable personal perception rather than noise or song-level confounds; the paper gives no inter-annotator reliability evidence or null baseline that would rule out MAML simply fitting label noise.","rationale":"The paper's central assertion is that DSAML predicts a new user's emotions from one annotated clip. For that to be true, per-annotator labels used to define tasks must contain a stable, transferable person-specific signal. The experiments provide some supporting evidence: Table 4 shows degradation when MAML or the annotator-based construction is removed, and the subjective ranking in Figure 6 is consistent with personalization. However, those comparisons are not sufficient. MAML could improve by fitting a support song's particular label trajectory and then predicting trajectories that are globally similar to that song; because support and query songs come from the same annotator, any systematic annotator-by-song assignment confound would masquerade as personalization. The absence of inter-annotator reliability statistics is especially important because DEAM is crowd-sourced and noisy; the paper itself says unstable annotations from 0-15s are removed, but it never quantifies how much stable person-specific signal remains. The risk is therefore not merely statistical: if the personalized effect is mostly noise, the PDMER formulation and the one-shot claim collapse. A permutation test over annotator identity directly falsifies or supports the assumption. Because the concern is substantial but testable, the appropriate verdict is CONDITIONAL, matching the reader's assessment.","tokens_in":11301,"tokens_out":8848,"duration_ms":96915,"concrete_test":"Run a permutation control for the task construction in Eq. (6): keep the exact training protocol, support/query sizes, and song identities, but randomly permute annotator labels across the 1000 training songs so that each pseudo-annotator receives labels from many true annotators. Train DSAML under this shuffled-annotator regime and evaluate on the same held-out personalized test tasks. If the shuffled model's arousal/valence CCC remains within a small margin (e.g., within 0.01) of Table 2, the annotator-based MAML gain is not attributable to stable user perception and the one-shot personalization claim fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The problem formulation and Eq. (6) make a load-bearing assumption: each annotator's labels are a learnable, transferable perceptual bias, so that one support sequence is enough to adapt to a new user. The paper never verifies this premise. DEAM is a crowd-annotated dataset, but the text does not report the number of annotators, their per-user song counts, or any inter-annotator agreement statistic such as ICC or per-user deviation relative to between-song variance. Without such statistics, the PDMER gains in Table 2 and the MAML/task-construction ablations in Table 4 can be explained by non-personalization mechanisms: (i) the inner loop overfits to the support song's idiosyncratic label trajectory, and (ii) annotator label deviations may correlate with song-level acoustic/emotional content because annotator-to-song assignment is not shown to be random. The claim of 'just one personalized annotation sample' is also undercut by an internal inconsistency: the inference description says all samples annotated by user p are used as support, while Implementation Details says Sp contains only one sample. At minimum, the one-sample adaptation claim cannot be validated without reporting exact test-task construction, per-user data statistics, and a null control that breaks annotator identity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes DSAML for dynamic music emotion recognition (DMER) and introduces a personalized variant (PDMER). The model uses a frozen ImageBind encoder with a trainable adapter to obtain dual-scale audio features, then applies a transformer with local and global attention masks plus a diagonal attention loss to capture short- and long-term context. Personalized prediction is achieved by formatting each annotator's labels as a meta-learning task and applying MAML with a task construction strategy that samples support and query sets from the same annotator. Experiments on DEAM and PMEmo report improvements over CRNN, DNN, MCRNN, and DAMFF in traditional DMER, and on DEAM in the personalized setting. A user study with 22 participants indicates that DSAML-generated curves are ranked significantly better than baseline curves. The paper makes a code and project page available.","tokens_in":11574,"tokens_out":6441,"duration_ms":60716,"significance":"If the results hold, this is a useful step for DMER: the dual-scale attention transformer with a diagonal attention loss is a sensible architectural contribution, and the annotator-based task construction is a natural way to formulate few-shot personalization for continuous emotion labels. The use of two public datasets, ablations covering each component, and a subjective evaluation are strengths, and the project page supports reproducibility. However, the significance of the PDMER claim depends on the assumption that per-annotator labels in DEAM reflect stable personal perception rather than noise; the manuscript does not currently provide evidence to distinguish these alternatives. The lack of statistical uncertainty and the small personalized test set further limit the strength of the state-of-the-art claims. With additional reliability analyses and a clearer evaluation protocol, the work could be an important contribution to personalized music emotion recognition.","major_comments":[{"comment":"The core PDMER claim assumes that each annotator's labels represent a stable personal perception that MAML can learn and transfer to a new user, but the paper provides no reliability evidence for DEAM per-annotator labels. Please report the number of annotators, per-annotator song counts, and an inter-annotator agreement statistic such as ICC or mean pairwise correlation on the 58 test songs. In addition, include a null control that breaks annotator identity (e.g., randomly permuting annotator IDs or replacing the support label with the mean label) and show that DSAML still outperforms it; otherwise the Table 2 gains and Table 4 ablations can be explained by fitting label noise or by non-random annotator-to-song assignment.","section":"Personalized Strategy, Eq. (6), and Objective Experiment, Table 2"},{"comment":"There is an internal inconsistency in the definition of the personalization support set. The inference paragraph says \"we use all samples annotated by a new user p ... in personalized data Sp\", while Implementation Details says \"both Si and Sp only contain 1 sample\"; the PDMER evaluation description also states that the test set uses personalized annotations from all annotators. Because the paper's headline claim is one-shot personalization, the exact protocol matters: how many test users, how many support clips per user, how query songs are chosen and whether they are disjoint from support songs, and whether labels in the test task are from a single annotator or aggregated. Please state the protocol explicitly and rerun the evaluation if the support set was in fact larger than one clip.","section":"Training & Inference Process and Implementation Details"},{"comment":"All objective results are reported as single point estimates without standard deviations, confidence intervals, or significance tests. Some improvements over baselines are small (e.g., PMEmo arousal CCC 0.186 vs 0.175 in Table 1, DEAM valence CCC 0.104 vs 0.055), and the personalized test set consists of only 58 songs. Please report mean and standard deviation over multiple training runs, and where possible paired tests across songs or users, so that the claimed state-of-the-art performance is statistically supported.","section":"Objective Experiment, Tables 1-4"},{"comment":"The subjective experiment is a key piece of evidence, but the reported paired t-test p-values are not accompanied by effect sizes or confidence intervals, and no correction for multiple comparisons is described. The statement that DSAML \"outperforms the ground truth\" in arousal should be clarified, since the comparison \"ground truth\" in Figure 6 is the dataset's mean annotator curve rather than a personalized ground-truth curve. Please report effect sizes, confidence intervals, and a precise definition of the ground-truth condition.","section":"Subjective Experiment, Figure 6"}],"minor_comments":[{"comment":"The notation Di = {(xi_1, yi_1), ...} reuses i both for the user index and the song index, which is confusing; please use a different subscript for the user.","section":"Problem Formulation, Eq. (6)"},{"comment":"The attention map panels are not described in the caption; please state which model variant each panel shows and how the diagonal attention loss changes the maps.","section":"Figure 4"},{"comment":"The sentence that the model without MAML but with the personalized task construction strategy is worse than the model without both components mostly matches the table, but the valence CCC is slightly higher (0.062 vs 0.061); please reword to be precise about the direction of the effect across metrics.","section":"Ablation Study, Table 4"},{"comment":"The claim that LSTM struggles with long-term dependencies is supported by NLP and time-series references; please clarify whether there is DMER-specific evidence or state that the transfer is a working assumption.","section":"Introduction and Related Work"}],"recommendation":"major_revision","confidential_remarks":"The main risk in this manuscript is that the PDMER evaluation may be overinterpreted: the test set is small (58 songs), the per-annotator reliability is unknown, and the one-sample support protocol is inconsistent between text sections. I would not reject on the basis of disagreement with the personalized-perception premise, but the authors need to provide reliability statistics and a null control to make the one-shot personalization claim convincing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe short version: this paper introduces a genuinely new problem—personalized dynamic music emotion recognition—and a sensible meta-learning strategy to tackle it. The \"divide tasks by annotator\" construction for MAML is the real contribution, and the dual-scale attention transformer with the diagonal attention loss is a reasonable engineering choice rather than a paradigm shift. The ablations show each component earns its place. Code and a demo site are promised, which helps.\n\nThe soft spots are mostly in the evaluation, and they are significant. The one-shot personalization claim rests on the assumption that per-annotator label differences are stable, transferable perceptual bias rather than noise or song-level confounds. The paper never checks this. No inter-annotator agreement statistic, no per-user song counts, no analysis of whether annotator-to-song assignment is random. Without that, the PDMER gains in Table 2 could come from the inner loop overfitting the support song's idiosyncratic label trajectory rather than from personalization. A null control that shuffles annotator identities would settle this, and it is absent.\n\nThere is also a real internal inconsistency: the inference text says \"all samples annotated by user p\" are used as support, while Implementation Details says Sp contains only one sample. That needs reconciling before the one-shot claim is credible.\n\nOn traditional DMER, the objective results lack error bars and significance tests, so the \"state-of-the-art\" claim is shaky. The paper itself concedes PMEmo valence PCC is below DAMFF, yet the abstract and conclusion still claim SOTA in both tasks—an overclaim. The subjective study is small (22 participants), though the paired t-tests are a reasonable attempt; effect sizes would be more informative.\n\nWho gets value: anyone working on music emotion recognition, especially personalization or few-shot adaptation. It is not a broad scientific reshape, but it is a solid, focused contribution. The paper deserves serious referee time. I would send it to review with the expectation of major revisions: add reliability analyses, a null baseline, confidence intervals, and tighten the claims.\n\nRecommendation: send to peer review, ask for those revisions before acceptance.","headline":"Sensible new problem and annotator-based MAML task construction, but the one-shot personalization claim outruns the evidence: no inter-annotator reliability check, no null baseline, no error bars, and a few overclaims.","tokens_in":12025,"tokens_out":2734,"would_cite":false,"duration_ms":24847,"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":"This paper claims that dynamic music emotion recognition can be made personal: a meta-learning model with annotator-specific tasks predicts a new listener's valence-arousal curve from one annotated clip, using a dual-scale attention…","keywords":["dynamic music emotion recognition","personalized music emotion recognition","meta-learning","MAML","dual-scale attention transformer","valence-arousal","few-shot adaptation","music information retrieval"],"falsifier":"Retrain DSAML on DEAM with annotator IDs randomly shuffled while keeping every clip-label pair identical, then run the same one-shot adaptation to held-out users; if shuffled-task performance matches real annotator-task performance, the personalization effect is an artifact, while a clear drop would confirm stable per-annotator perception. A second concrete check is to vary the support set from one clip to two or three of the user's own annotations: genuine personalization should improve with more user-specific data, whereas pure averaging or architecture gains would stay flat.","tokens_in":11141,"feed_emoji":"🎵","tokens_out":10477,"duration_ms":102694,"temperature":0.7,"pith_summary":"This paper tries to establish that music emotion recognition should predict the emotion curve of a specific listener rather than the averaged emotion of a crowd, and that this can be done with only one annotated clip from that listener. It proposes DSAML, a model with a dual-scale feature extractor and a dual-scale attention transformer that handles both short, local emotional changes and the long-range arc of a piece. For personalization it uses MAML meta-learning with tasks built from individual annotators' labels, so each task corresponds to one person's perception and the model learns to adapt quickly to any new person. The reported experiments say the model beats earlier LSTM-based methods on the group-average task on DEAM and PMEmo, and, in both objective and subjective tests, predicts personalized V-A curves better than baselines. If the claim holds, personalized emotional music recommendation and music therapy could work from a single example of a user's reaction.","feed_headline":"One annotated clip is enough to personalize music emotion prediction","feed_subtitle":"Dual-scale attention and annotator-level meta-learning beat group-mean baselines, and one clip is enough to personalize.","key_machinery":"The engine is the dual-scale attention transformer: one transformer with shared weights runs twice, once under a local mask with context length $n_l$ and once under a global mask with context length $n_g$, where $n_g \\gg n_l$. The diagonal attention loss $L_{\\text{attention}} = \\tfrac{1}{n}\\sum_i [(\\operatorname{diag}(A_l)_i - \\alpha)^2 + (\\operatorname{diag}(A_g)_i - \\beta)^2]$ with large $\\alpha$ and small $\\beta$ forces the local attention map to stay near the diagonal and the global map to spread out, so the two views do not collapse into the same features. On the personalization side, the load-bearing mechanism is the annotator-based task construction inside MAML: each task's support and query sets are clips labeled by the same annotator, making each task one personality, and the meta-initialization learned across those tasks can be updated from one clip from a new annotator.","core_discovery":"The core claim is that a listener's emotional annotations are not noise to be averaged away but signal to be learned, and that a meta-learning setup with one task per annotator can turn that signal into a one-shot personalization mechanism. DSAML's dual-scale attention transformer applies two context masks to one shared transformer, a local context and a global context, and adds a diagonal attention loss so the two scales genuinely separate, letting the model use both moment-level detail and the overall emotional trajectory. On the conventional DMER setting the paper reports it outperforms prior baselines on DEAM and PMEmo; on the PDMER setting it reports that adapting the meta-learned model to an unseen annotator from a single support clip improves prediction of that annotator's own V-A curve, with a subjective listening test placing DSAML ahead of the averaged-label ground truth for arousal.","pith_inferences":["Beyond the paper, the annotator-as-task recipe applies to any domain where raters disagree systematically, such as perceived stress or video valence, turning rater disagreement into a source of training tasks rather than noise.","The one-shot claim would be stronger if tested across repeated listening sessions; listener perception may drift over time, which would require adapting the meta-initialization per session rather than once per user.","A natural untested extension is to predict a personalized uncertainty or confidence band for the valence-arousal curve, since listeners who annotate less consistently should be adapted less aggressively.","The architecture's long-range mask could be combined with a shorter context length on the global branch for very long tracks, trading some context for computation while preserving the dual-scale behavior."],"forward_implications":["A new listener can be onboarded with a single annotated clip, because the meta-learned initialization already encodes how emotional perception varies across people.","Datasets that keep annotator IDs become much more useful: the same clips generate one meta-task per annotator instead of one averaged label, multiplying the available training signal.","The dual-scale mask-and-loss pattern is not music-specific and should carry over to other continuous emotion regression tasks, such as video, speech, or physiological signals, where local and global context both matter.","A single DSAML model can serve both traditional group-level prediction and one-shot personalization, since the same transformer and meta-learner produce both averaged and personalized outputs."],"supporting_citations":[{"why":"Supplies the DEAM dataset with per-annotator dynamic valence-arousal annotations, which both the personalized task construction and the personalized test set depend on.","marker":"[Aljanaki, Yang, and Soleymani 2017]"},{"why":"Supplies the MAML meta-learning update that DSAML uses to fast-adapt to a new annotator from one support clip.","marker":"[Finn, Abbeel, and Levine 2017]"},{"why":"Supplies the frozen pretrained audio feature model whose global embeddings are fused with the adapter's local features.","marker":"[Girdhar et al. 2023]"},{"why":"Provides the MCRNN BiLSTM baseline and the multi-view CNN plus BiLSTM design that DSAML improves on in the long-term dependency comparison.","marker":"[He and Ferguson 2020]"},{"why":"Provides the DAMFF state-of-the-art DMER baseline and the BiLSTM-based multi-scale fusion approach that DSAML directly compares against.","marker":"[Zhang et al. 2023]"},{"why":"Presents evidence that recurrent sequence models use local context and struggle with long-range dependencies, motivating the dual-scale attention transformer.","marker":"[Khandelwal et al. 2018]"},{"why":"Documents that earlier personalized music emotion recognition required at least 20 user-annotated samples, establishing the motivation for one-shot PDMER.","marker":"[Chen et al. 2014]"}],"fun_headline_variants":["One-shot personalization for music emotion recognition","Your music mood from one clip: meta-learning wins","Dual-scale attention turns one annotation into personalized emotion","Personalized V-A curves with a single support clip","Beat the average: one clip personalizes music emotion AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire personalized pipeline rests on one assumption: the different labels different annotators give to the same song reflect stable, learnable differences in emotional perception rather than annotation noise, because if they are mostly noise, adapting to a new user from one annotated clip cannot produce genuine personalization.","fun_headline_variants_meta":{"raw":{"variants":["One-shot personalization for music emotion recognition","Your music mood from one clip: meta-learning wins","Dual-scale attention turns one annotation into personalized emotion","Personalized V-A curves with a single support clip","Beat the average: one clip personalizes music emotion AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000359,"raw_usage":{"total_tokens":1944,"prompt_tokens":945,"completion_tokens":999,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":924}},"tokens_in":561,"tokens_out":999,"duration_ms":9219,"temperature":1.0,"reasoning_tokens":924,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:49:48.961526+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain DSAML on DEAM with annotator IDs randomly shuffled while keeping every clip-label pair identical, then run the same one-shot adaptation to held-out users; if shuffled-task performance matches real annotator-task performance, the personalization effect is an artifact, while a clear drop would confirm stable per-annotator perception. A second concrete check is to vary the support set from one clip to two or three of the user's own annotations: genuine personalization should improve with more user-specific data, whereas pure averaging or architecture gains would stay flat.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the DEAM dataset with per-annotator dynamic valence-arousal annotations, which both the personalized task construction and the personalized test set depend on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the DAMFF state-of-the-art DMER baseline and the BiLSTM-based multi-scale fusion approach that DSAML directly compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents that earlier personalized music emotion recognition required at least 20 user-annotated samples, establishing the motivation for one-shot PDMER."}],"review_version":1}