{"id":"00ec885e-df0e-4064-8667-f9d9b0ecd2a8","arxiv_id":"2504.17834","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"GUSD claims state-of-the-art spoiler detection on IMDb datasets using genre-aware routing and user-bias features from dynamic graph pretraining.","lead":"A new model, GUSD, detects spoilers in movie reviews by adding movie genre and reviewer history signals to a graph neural network. On two IMDb review benchmarks it beats previous methods by large margins, though the evaluation setup may inflate the gains.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Random review split lets GUSD exploit the same users' labeled history in test, so its reported user-bias advantage may not hold for unseen users; a user-disjoint split is the decisive check.","rationale":"The reader's weakest-assumption analysis is correct and is the most load-bearing issue I found. The paper's central claim is SOTA performance, and the user-bias mechanism is a key novelty. The random split lets test reviews inherit the spoiler-posting history of users who also have labeled training reviews, which is a strong and legitimate feature in deployment only if the platform has that history. The paper never tests the cold-user case, so we cannot tell how much of the gains come from the proposed dynamic-graph user bias versus from simple user-history leakage across the split. The ablation removing U_b is not decisive because the graph still carries user identity and history. A user-disjoint split directly probes the generalization of the central component. I also note the LM mismatch (GUSD uses Bge-Large on Kaggle while MVSD uses RoBERTa) and the missing significance tests, but those are secondary; the user split is the first-order validity check. Since my concern matches the reader's identified weakest assumption, the verdict remains CONDITIONAL, pending the proposed experiment.","tokens_in":14324,"tokens_out":5632,"duration_ms":60016,"concrete_test":"Construct a user-disjoint split of the Kaggle dataset: assign all reviews by each user to the same split, ensuring test users have zero reviews in the training set, while preserving the overall 70/20/10 ratio and the spoiler rate per split as closely as possible. Retrain GUSD and MVSD (and ideally SpoilerNet and R-GCN) on this split with otherwise identical settings, and compare F1/AUC/Acc. If GUSD's margin over MVSD shrinks substantially (e.g., by more than half of the original 15.2-point F1 gap), the reported state-of-the-art result is largely attributable to the random-split user overlap rather than to the proposed user-bias modeling.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is state-of-the-art spoiler detection, and the paper's newest component is user bias obtained from dynamic graph modeling of review history. Under the 7:2:1 random review split (Sec. 4.1), the same users appear in training and test, so a test review's user node has a history of labeled training reviews. The dynamic graph pretraining and the downstream R2GFormer message passing can therefore propagate that user's observed spoiler rate to the test review, effectively making user identity a strong test-time feature. The paper's own ablation (Table 2, 'w/o U_b') shows removing the explicit user-bias vector costs 1.4 F1, but this understates the issue because the graph still connects test reviews to user nodes with labeled training edges; in a user-disjoint split that entire signal disappears. The paper never evaluates this cold-user scenario, although the model's motivation and the case study (Table 4) both emphasize user history. For the SOTA claim to hold as stated, GUSD's advantage must persist when user history is unavailable at test time, or the claim must be conditioned on platforms with known user histories. The risk is that most of the 15.2-point Kaggle gain relative to MVSD comes from access to overlapping user histories rather than from the genre-aware architecture.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GUSD, a spoiler detection framework that combines genre-aware modules (GenreFormer and GMoE) with user bias extracted from dynamic graph modeling of review history. The framework is evaluated on the LCS and Kaggle datasets against 14 baselines. GUSD reports state-of-the-art results, improving binary F1 by 15.2 points on Kaggle and 6.1 points on LCS over the previous best method MVSD. The paper provides ablations, robustness studies, and case studies, and it releases source code.","tokens_in":14579,"tokens_out":6394,"duration_ms":57367,"significance":"If the reported results are robust, GUSD would be a meaningful advance: it is the first spoiler detector to combine dynamic-graph user bias with genre-aware mixture-of-experts routing, and the reported gains over strong graph baselines are large. The paper's strengths include releasing code, running five runs and reporting standard deviations, and providing ablations. However, the central SOTA claim is currently weakened by two evaluation confounds: the random review split lets the model exploit overlapping user histories at test time, and on Kaggle GUSD uses a stronger text encoder (Bge-Large) than the GNN baselines (RoBERTa). The magnitude of the reported gains therefore cannot be attributed to the proposed architecture alone without additional controlled experiments.","major_comments":[{"comment":"Section 4.1 states that reviews are randomly split 7:2:1. Because users and movies appear in both training and test, the graph built in Section 3.2 (edges E2/E3) connects each test review to a user node that also has edges to that user's labeled training reviews; the user-bias vector from dynamic graph pretraining (Section 3.1) is likewise estimated from the same user's review history. The classifier can therefore use the user's observed spoiler rate as a test-time feature. The ablation in Table 2 (w/o U_b) removes only the explicit bias vector, not the graph connections, so it does not estimate performance for unseen users. Please report results on a user-disjoint split (and ideally a movie-disjoint split), or explicitly restrict the SOTA claim to platforms with known user histories.","section":"Sec. 4.1 / Sec. 3.2"},{"comment":"Table 5 shows that on Kaggle GUSD uses Bge-Large as its text encoder, while the GNN baselines (including MVSD) use RoBERTa; the text-encoder gap is a confound for the 15.2-point F1 improvement in Table 1. Please add a controlled comparison, e.g., GUSD with RoBERTa on Kaggle or MVSD with Bge-Large, to separate the contribution of the proposed architecture from that of the encoder.","section":"Table 5 / Table 1"},{"comment":"Section 3.1 says user bias is obtained by converting the static dataset into a dynamic event stream and pretraining DyGFormer with link prediction, but neither the event-stream construction nor the link-prediction training (negative sampling, temporal split, loss) is described, so the user-bias component is not reproducible from the paper. The supplementary material also does not provide these details.","section":"Sec. 3.1 / Appendix"}],"minor_comments":[{"comment":"The reported gains of 15.2% and 6.1% are percentage-point differences, not relative percentages; please correct the wording.","section":"Sec. 4.2"},{"comment":"The notation overloads the symbols U, R, and M: the left-hand sides are node representations after a TRM, while the right-hand sides redefine the same symbols with an MLP; please use distinct notation.","section":"Sec. 3.3, Eq. (9)-(11)"},{"comment":"Table 6 lists 'genres' as unavailable for LCS, which seems to contradict Section 4.1's statement that both datasets include genre information; please clarify how genres were obtained for LCS.","section":"Table 6 / Sec. 4.1"},{"comment":"The claim of statistical significance in Section 4.2 is not backed by any significance test; please report confidence intervals or paired tests.","section":"Sec. 4.2"},{"comment":"Minor typos: 'dervied' in Eq. (11) and 'Catagory' in the Table 2 header.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is competently written and the proposed framework is interesting, but the two evaluation confounds (user overlap in the random split and the asymmetric text encoders on Kaggle) need to be resolved before the SOTA claim can be accepted. Please ask the authors to provide controlled experiments, especially a user-disjoint split and an LM-matched comparison. The novelty claim of being 'first to model' is somewhat overstated given SpoilerNet and DNSD, but that is a presentation issue rather than a technical flaw."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The first thing to know: this paper proposes a new spoiler-detection framework, GUSD, that combines dynamic-graph user embeddings, a k-hop graph attention layer (RetGAT), genre-conditioned transformer pooling (GenreFormer), and a genre-routed mixture of experts (GMoE). It reports large state-of-the-art gains over the previous best MVSD on two benchmarks. The code is public, and the architecture is described in detail. The real question is whether the reported gains come from the model or from the experimental setup.\n\nWhat's genuinely new is the packaging: nobody has modeled user spoiler tendency as a dynamic-graph embedding and combined it with explicit genre routing for spoiler detection. The paper also presents a useful corrected algorithm for computing k-hop neighbors, and the genre-routing idea is a sensible way to use metadata that is usually ignored. The ablations mostly support the value of each component, and the robustness tests are a nice addition.\n\nNow the soft spots. On Kaggle, GUSD uses Bge-Large as its text encoder, while the GNN baselines use RoBERTa. That makes the 15.2-point F1 gain over MVSD very hard to attribute to the architecture. A matched-LM comparison is missing. On top of that, the random 7:2:1 split means the same users appear in training and test; user bias is computed from review history, so for a test review the model has access to that user's labeled spoiler behavior. The ablation that removes the explicit user-bias vector costs only 1.4 F1, but the graph edges still connect test reviews to user nodes with training history, so the leak may persist. The paper never tests the cold-user scenario. The evaluation is also transductive, which is not stated. These issues together mean the SOTA claim is not yet supported.\n\nThe architecture is complex, but the writing is clear and the code release helps. I would send this to peer review because it is a serious, reproducible piece of work with a real gap in evaluation; the reviewers should ask for a matched-LM baseline, a user-disjoint split, and ideally error bars on the ablations.\n\nFor a reader, this is most useful if you are working on spoiler detection or content moderation with user metadata. The techniques themselves are not going to change the field, but they are a reasonable engineering contribution.","headline":"A serious, code-released spoiler-detection architecture with plausible components, but the SOTA claim is undercut by a stronger LM on Kaggle and a leaky random split that lets user history boost test performance.","tokens_in":15142,"tokens_out":3658,"would_cite":false,"duration_ms":35361,"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 adding a reviewer's past spoiler behavior and the movie's genre to a graph-based detector lifts spoiler-detection F1 by 15.2 points on Kaggle and 6.1 points on LCS over the previous best method.","keywords":["spoiler detection","movie genre","user bias","dynamic graph","mixture of experts","graph attention network","IMDb reviews","review classification"],"falsifier":"Run GUSD with a user-disjoint split in which all reviews by a given user are kept in the same fold, and compare its F1 against MVSD. If the gap narrows to near zero or reverses, the user-bias component is exploiting review-history label leakage rather than a transferable user trait; if the gap persists, the user signal generalizes to unseen users.","tokens_in":14117,"feed_emoji":"🎬","tokens_out":8269,"duration_ms":68563,"temperature":0.7,"pith_summary":"The paper argues that spoiler detection in movie reviews has been missing two cheap signals: the genre of the movie being reviewed and the reviewer's own tendency to post spoilers. It reports that spoiler rates differ strongly by genre—plot-driven genres like Film-Noir and Adventure have more spoilers—and that a small subset of users accounts for a disproportionate share of spoiler posts. To exploit this, the paper introduces GUSD, which learns a per-user bias from a dynamic graph of review history, runs a genre-aware graph attention layer over users, reviews, and movies, and routes each review through genre-specific expert classifiers. On the standard IMDb-based benchmarks, GUSD lifts binary F1 to 80.24 on Kaggle and 75.37 on LCS, beating the previous best method by 15.2 and 6.1 points respectively. If the result holds, genre and user history are not side information but core features for content moderation on review platforms.","feed_headline":"Spoiler detection F1 jumps 15 points with genre and user bias","feed_subtitle":"GUSD reads a reviewer's spoiler history and the film's genre, beating the old best by 15.2 F1 points.","key_machinery":"The load-bearing mechanism is GUSD's three-way fusion of graph, genre, and user-history signals. RetGAT aggregates node features inside k hops with a decay factor $\\delta_h = \\exp(-\\alpha h)$, so nearby users and movies influence a review more strongly than distant nodes while distant structure is truncated. GenreFormer pools all review and movie nodes of a genre into a genre vector, exchanges those vectors across genres with a transformer, then attaches the averaged genre vector to each node via concatenation and an MLP. GMoE routes every review to the expert MLPs of its movie's genres, bypassing learned gating in favor of explicit genre assignment. User bias comes from pre-training a dynamic graph encoder on the temporal sequence of reviews and using the user node's embedding as a feature; in the final layer the review's graph embedding, its metadata, and the user-bias vector are concatenated and classified.","core_discovery":"The central claim is that jointly modeling genre and user-specific spoiler propensity gives a large, previously unreached jump in spoiler-detection accuracy. The paper shows that spoiler prevalence varies by genre, with plot-heavy genres more often spoiled, and that users differ widely in spoiler rates. GUSD operationalizes this with three components: RetGAT, a graph attention network that aggregates features over k-hop neighborhoods with exponential decay; GenreFormer, which builds a per-genre representation from all review and movie nodes of that genre and lets genres exchange information through a transformer; and GMoE, a genre-conditional mixture of experts that assigns each review to the experts of the genres its movie belongs to. User bias is pre-trained separately by encoding the stream of user-review-movie events with a dynamic graph encoder and keeping the resulting user embedding. Fusing these signals, GUSD reports a binary F1 of 80.24 on Kaggle and 75.37 on LCS, exceeding MVSD, the previous best, by 15.2 and 6.1 points, with the ablation studies attributing the gain to the user-bias vector and genre routing rather than to the graph backbone alone.","pith_inferences":["The random 7:2:1 split means each user appears in training before test, so the user-bias embedding can read that user's labeled history; a user-disjoint split would likely shrink the reported gain and is the right next experiment.","The same user-history graph could generalize to other per-user moderation tasks, such as detecting spam or harassment, where posting history is a stable trait; that extension is untested here.","If genre routing is the real driver, then using a lightweight genre classifier at deployment could make GUSD work on platforms without explicit genre tags; the paper does not evaluate that transfer.","The authors' corrected k-hop neighbor computation may matter beyond this paper, since prior multi-hop GNNs relied on the matrix-power approximation; an isolated ablation would show how much of the gain comes from that correction alone."],"forward_implications":["Spoiler-detection services can improve accuracy without new text models by feeding in the movie's genre tags and the reviewer's past spoiler rate.","Dynamic graph encoders trained on review streams can extract a reusable user trait that predicts spoiler posting on unseen reviews of the same user.","Explicit genre-based routing of reviews to specialized classifiers beats learned routing (standard MoE and Soft-MoE) on this task, suggesting that hard domain labels are useful for expert assignment.","The reported F1 of 80.24 on Kaggle and 75.37 on LCS sets a new reference point for future spoiler-detection work, and the released code allows direct comparison."],"supporting_citations":[{"why":"Supplies the LCS dataset, the random split, and the MVSD baseline that GUSD must exceed.","marker":"[38]"},{"why":"DNSD is the prior genre-aware attention baseline that GUSD outperforms, demonstrating the insufficiency of simple genre concatenation.","marker":"[6]"},{"why":"SpoilerNet is the prior user/item-bias baseline; comparing against it isolates the benefit of dynamic-graph user modeling.","marker":"[36]"},{"why":"Provides the dynamic graph encoder whose user node embeddings become GUSD's user-bias feature.","marker":"[42]"},{"why":"The sparsely-gated mixture-of-experts layer is the baseline that GMoE replaces in the ablation study.","marker":"[33]"},{"why":"Soft-MoE is the other contemporary mixture-of-experts baseline compared in the GMoE study.","marker":"[29]"},{"why":"RoBERTa is the text encoder used for the LCS dataset and one of the pretrained-language-model baselines.","marker":"[19]"},{"why":"The Kaggle IMDb spoiler dataset is one of the two benchmark datasets used for the main results.","marker":"[24]"}],"fun_headline_variants":["Spoiler detection F1 jumps 15 points via genre and user bias","Genre and user history lift spoiler detection by 15 F1","GUSD: genre-aware spoiler detection gains 15 points","15-point spoiler detection boost from genre and user patterns","User bias and genre: hidden keys to spoiler detection leap"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation splits reviews randomly (7:2:1), so the same users and movies appear in training and test; the user-bias vector for a test review is built from that user's labeled training reviews, and for a brand-new user with no history the bias signal would be unavailable, which would likely shrink GUSD's advantage over text-only methods.","fun_headline_variants_meta":{"raw":{"variants":["Spoiler detection F1 jumps 15 points via genre and user bias","Genre and user history lift spoiler detection by 15 F1","GUSD: genre-aware spoiler detection gains 15 points","15-point spoiler detection boost from genre and user patterns","User bias and genre: hidden keys to spoiler detection leap"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000289,"raw_usage":{"total_tokens":1735,"prompt_tokens":1029,"completion_tokens":706,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":645,"completion_tokens_details":{"reasoning_tokens":617}},"tokens_in":645,"tokens_out":706,"duration_ms":6246,"temperature":1.0,"reasoning_tokens":617,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:33:55.330969+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run GUSD with a user-disjoint split in which all reviews by a given user are kept in the same fold, and compare its F1 against MVSD. If the gap narrows to near zero or reverses, the user-bias component is exploiting review-history label leakage rather than a transferable user trait; if the gap persists, the user signal generalizes to unseen users.","supporting_citations":[{"cited_title":"In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing","cited_arxiv_id":null,"evidence_quote":"Supplies the LCS dataset, the random split, and the MVSD baseline that GUSD must exceed."},{"cited_title":"In: Advances in Knowledge Dis- covery and Data Mining: 22nd Pacific-Asia Conference, PAKDD 2018, Melbourne, VIC, Australia, June 3-6, 2018, Proceedings, Part I 22","cited_arxiv_id":null,"evidence_quote":"DNSD is the prior genre-aware attention baseline that GUSD outperforms, demonstrating the insufficiency of simple genre concatenation."},{"cited_title":"Key Information","cited_arxiv_id":null,"evidence_quote":"Provides the dynamic graph encoder whose user node embeddings become GUSD's user-bias feature."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The Kaggle IMDb spoiler dataset is one of the two benchmark datasets used for the main results."}],"review_version":1}