{"id":"4fce1691-5711-4c38-af63-88bed5be1175","arxiv_id":"2608.11801","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"JAPE is a framework that forecasts anomalies in multivariate time series from evolving inter-variable dependency graphs, and reuses those graphs to explain which variables drive each alert.","lead":"JAPE predicts future anomalies in multivariate time series by modeling how relationships between variables change, not just how the recorded values change. It also ranks which variables are most involved in a predicted anomaly using the same dependency graphs, at no extra computation cost.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Learned dependency graphs are not validated as true structural precursors; the reported gains may stem from the forecasting backbone and supervised alerting head rather than dependency modeling.","rationale":"I read the paper as proposing that dynamic directed dependency graphs are the key novelty; the empirical evidence for this is the ablation in Table 4 and the explanation comparison in Table 3. Both are single-run and lack error bars, but more importantly, neither verifies that the learned graphs encode the dependency changes the paper claims. The graph construction is entirely unsupervised with respect to structure, and the reported sensitivity to K_max and K_g shows the graph can be unstable. The WADI drop is direct evidence that the graph can hurt. A synthetic test with known ground truth would settle whether the graph is meaningful or a spurious byproduct of forecasting. I agree with the reader's conditional verdict; the paper should not be accepted as-is without this validation or comparable real-world evidence (e.g., alignment with known causal annotations).","tokens_in":18444,"tokens_out":6848,"duration_ms":73406,"concrete_test":"Generate a synthetic VAR(2) process with known lagged edges and inject anomalies as dependency changes (e.g., an edge disappearing or reversing) while keeping numerical distributions similar. Train JAPE, measure edge-recovery F1 of the learned top-K graph against ground truth, and compare alerting and explanation performance using true graphs, learned graphs, and randomly permuted graphs with identical degree sequence. If edge-recovery F1 is near chance, or if permuting edge identities leaves F1/AUC-PR/MRR unchanged, the structural signal is not edge-specific and the central dependency-modeling claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the learned top-K lag-aware directed graphs (Eq. 3-4) capture dependency changes that precede anomalies. The graph module is trained only through the forecasting MSE (Section 3.1.4) and the alerting head's focal loss; no structural supervision or validation ensures the selected edges correspond to true lead-lag relationships. Since S_uv = q_u·ek_v − k_u·eq_v is an unnormalized inner-product contrast, it can be dominated by scale and synchrony rather than directed predictive influence, and TopK with K_g=5 plus row-normalization produces a sparse, row-stochastic graph whose edge identities may be arbitrary. The alerting and explanation modules inherit this graph, so if it is noisy the reported gains are not attributable to dependency-structure modeling. The paper's own WADI ablation (Section 4.4) shows a 1.0 F1 drop when the graph is used, and Appendix E shows K_max=5 collapses MSL F1 from 61.2 to 35.5, demonstrating that the learned graph can add substantial noise. Without direct evidence of graph validity, the 19.7% F1 / 41.3% AUC-PR gains and the 26.6% MRR explanation gain may be driven by the numerical forecasting backbone and supervised alerting head rather than by the dynamic dependency structure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes JAPE, a framework for multivariate time-series anomaly prediction over a future horizon H, together with variable-level explanations at alert time. The Decoupled Spatio-Temporal Representation (DSTR) backbone produces both point forecasts and dynamic directed dependency graphs through a lag-aware directional contrast score and top-K graph construction (Eqs. 1-4). A dual-view alerting head fuses forecast tokens with structural descriptors derived from the graph sequence, and the Native Predictive Explanation (NPE) module reuses the predicted graphs to compute a Graph Deviation Score that ranks variables. Experiments on SMD, WADI, MSL, PSM, and EXATHLON report large average gains in F1, AUC-PR, and MRR over the tested baselines.","tokens_in":18733,"tokens_out":5561,"duration_ms":58204,"significance":"If the reported results hold, the paper makes a useful contribution: it is, to my knowledge, the first anomaly-prediction method to use evolving dependency graphs for both point-wise alerting and native variable-level explanation, and the design is described in enough detail to reimplement, with code provided. The use of strict point-wise evaluation and the inclusion of ablations for the backbone and the graph signal are strengths. The main caveat is that the quantitative claims currently rest on single runs and on a graph module whose quality is not directly validated; the WADI ablation and the K_max sensitivity results show that the graph can also be a source of noise. The central idea is defensible and the manuscript is worth publishing after the evaluation is strengthened.","major_comments":[{"comment":"The ablation labeled 'w/o Dynamic dependency graph' removes the graph input from the alerting head, but it does not compare against a static-graph variant or against an existing graph-based detector such as GDN or MTAD-GAT. As a result, the reported average gains cannot be attributed specifically to evolving dependency modeling rather than to the DSTR backbone or the supervised alerting head. Please add a static-versus-dynamic graph comparison or a graph-based baseline that shares the same forecasting backbone and alerting head.","section":"Section 4.4, Table 4"},{"comment":"All reported numbers are single runs, with no error bars, standard deviations, or significance tests. Given that Table 6 shows K_max=5 collapses MSL F1 from 61.2 to 35.5 and K_g choices swing results by several F1 points on SMD and MSL, the headline improvements of 19.7% F1 and 41.3% AUC-PR could be within hyperparameter or seed noise. Please report multiple seeds with mean and standard deviation, and add a significance test or at least per-seed results for the main comparisons.","section":"Section 4.2, Tables 2, 6, and 7"},{"comment":"The NPE explanation scores are computed from graphs that are never validated as true structural precursors. The graph is trained only through forecast MSE and the alerting head's focal loss (Algorithm 1, Stages 1-2), so nothing ensures that the top-K edges correspond to real lead-lag relationships. The WADI result in Table 4 and the K_max sensitivity in Table 6 show that the learned graph can add noise. Please add (i) a sanity check of learned edges against known or plausible dependencies, for example using a synthetic dataset with ground-truth structure or the process annotations in SMD, and (ii) a graph-agnostic attribution baseline so that the value of the graph can be separated from the value of the forecasting representation.","section":"Section 3.3 and Section 4.3"},{"comment":"The paper motivates JAPE with 'a practical setting where a small number of anomaly labels are available,' but no experiment varies the number of labeled windows; all supervised methods use the full label set. Add an ablation that scales the fraction of labeled windows used for alert-head training to support this practical claim.","section":"Introduction and Section 4.2"}],"minor_comments":[{"comment":"The sentence immediately after Eq. (7) contains a duplicated phrase: 'Specifically, Specifically,'. Please remove the duplicate.","section":"Section 3.3"},{"comment":"The text contains missing spaces from LaTeX macros, e.g., 'JAPEis' in the abstract and 'aJoint' in the introduction. Please fix the rendering throughout.","section":"Abstract and Introduction"},{"comment":"Several entries are hard to read because adjacent numbers are not separated, e.g., '53.8 54.642.562.8' and '51.544.9'. Reformat the table so each (F1, AUC-PR) pair is clearly delimited.","section":"Table 2"},{"comment":"The explanation evaluation reports only point estimates on SMD and WADI, and WADI contains only 15 events. Please add at least standard deviations or per-event breakdowns for the explanation metrics, and state how the normal reference windows for GDS are selected (size and position).","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is publishable after major revision. The 'first' novelty claim should be checked carefully against MultiRC, F2A, and graph-based anomaly-detection precursors; the related-work discussion separates these lines but does not explicitly rule out an existing method that jointly predicts and explains via dependency structures. Also, several references to trajectory-compression papers by the same group (e.g., [10], [18]-[20], [40]) are not directly relevant to the anomaly-prediction contribution and could be trimmed; this is a presentation concern rather than a scientific one."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: JAPE is a clean, implementable combination of dynamic directed graph learning, point-wise future anomaly prediction, and a zero-cost explanation module that reuses the predicted graph. The central claim—that modeling evolving inter-variable dependencies adds signal beyond numerical forecasting—is plausible and mostly supported by the ablations. There is no fatal flaw, but the empirical support is thinner than the 19.7% F1 / 41.3% AUC-PR headline suggests.\n\nWhat's genuinely new: the decoupled spatio-temporal backbone with lag-aware directional contrast, the dual-view alerting that fuses numerical and structural evidence, and the graph deviation score for native variable-level explanation. The combination appears new relative to the cited work, which is either value-centric prediction or graph-based detection. The evaluation is strict (point-wise matching, no adjustment), and the method is specified well enough to reimplement. The ablation that removes the dynamic graph is the right test, and it shows a real average gain (62.7 to 65.0 F1), though with a 1.0 F1 drop on WADI.\n\nSoft spots, in proportion: no error bars or significance tests anywhere, so we don't know if the gains are stable across runs. No graph-based baseline is compared—GDN, MTAD-GAT, or even a random-edge version of the same graph would have shown whether the structure itself matters or just extra parameters. The explanation evaluation uses only two datasets, one with just 15 events. The 'small number of anomaly labels' setting is claimed but never ablated. The stress-test note about graph validity is fair: the learned edges are only supervised through forecasting MSE and the alerting head, so they could be noisy, and the sensitivity analysis shows that a bad lag range can collapse MSL F1 from 61.2 to 35.5. That is a caution, not a fatal flaw, but it should be addressed directly in revision.\n\nWho this is for: anyone working on time-series anomaly prediction or explainable multivariate monitoring. It is a useful, honest piece of work with clear limitations that are partly acknowledged. It deserves a serious referee, not a desk reject, but it needs revision: error bars, a graph-based baseline or random-edge control, and a more thorough explanation evaluation. I would take it for peer review and expect heavy revision.","headline":"A coherent, well-specified framework that makes a plausible case for dependency-structure modeling in anomaly prediction; the headline numbers are stronger than the evidence, but the paper deserves a serious referee.","tokens_in":19284,"tokens_out":1737,"would_cite":true,"duration_ms":20234,"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":"JAPE claims that forecasting evolving directed dependency graphs alongside numerical values improves multivariate time-series anomaly prediction and yields variable-level explanations with no extra model.","keywords":["time series anomaly prediction","multivariate time series","dynamic dependency modeling","variable-level explanation","point-wise alerting","lag-aware graph learning","decoupled spatio-temporal representation"],"falsifier":"Run a permutation control on SMD, MSL, and PSM where, at inference, the learned source variables in each graph slice are randomly rewired before the alerting head and the Graph Deviation Score are applied; if F1, AUC-PR, and MRR do not clearly drop relative to the true graphs, then the dependency structure is not carrying the reported gains.","tokens_in":18209,"feed_emoji":"🔗","tokens_out":5155,"duration_ms":55282,"temperature":0.7,"pith_summary":"JAPE sets out to show that multivariate time-series anomalies are better predicted from changes in how variables depend on each other than from deviations in future numerical values alone. The paper argues that weak anomaly precursors often appear first as rearrangements of directed inter-variable dependencies, so it proposes a framework that forecasts both future values and a sequence of dynamic dependency graphs, fuses the two views for point-wise alerting, and reuses the graphs to rank the variables most involved in each alert. If this is right, operators can get earlier alarms and an intrinsic variable-level explanation from the same forward pass, without a separate post-hoc attribution model. On five real-world benchmarks across three horizons the reported result is an average 19.7% gain in F1, 41.3% in AUC-PR, and 26.6% in MRR over the strongest baselines under strict point-wise evaluation.","feed_headline":"Modeling variable links, not just signal values, forecasts anomalies","feed_subtitle":"JAPE gains 19.7% F1 and 41.3% AUC-PR on five benchmarks, and explains alerts without a separate model.","key_machinery":"The load-bearing object is the lag-aware directed dependency graph $\\mathbf{A}_i$, built from a directional contrast score $\\mathbf{S}_i = \\mathbf{Q}_i(\\tilde{\\mathbf{K}}_i)^\\top - \\mathbf{K}_i(\\tilde{\\mathbf{Q}}_i)^\\top$: for an ordered pair $(u,v)$, a positive score indicates that the past of $v$ helps predict the present of $u$, and a ReLU-tanh Top-$K$ gate with $K_g=5$ turns those scores into sparse directed edges per target variable. This graph is produced inside a decoupled spatio-temporal backbone whose temporal axis is channel-independent and whose spatial axis aggregates lagged projections with learnable exponential decay, plus a future-oriented branch that extends graph construction into the forecast horizon. The same graph sequence is compressed into 13-dimensional structural descriptors for the dual-view alerting head, and it is directly reused by the Graph Deviation Score for variable-level explanation. The machinery's work is to make structural precursors explicit, reusable, and trainable under a pure forecasting objective before the alerting head is optimized.","core_discovery":"The central claim is that modeling evolving dependency structure is not merely an auxiliary signal but a primary carrier of anomaly-predictive evidence. JAPE constructs, for each patch position and encoder layer, a lag-aware directed graph representing directional influence from the past of one variable to the present of another, and it builds these graphs across both the historical window and the prediction horizon. Alerting then fuses a numerical forecast view with a structural view through cross-attention, so structural evidence can contribute even when the source variable's numerical deviation is subtle. Explanation reuses the very same predicted graphs: a Graph Deviation Score compares the alerted segment's aggregated graph against a normal reference graph, adding direct and multi-hop dependency deviations, and ranks variables at negligible extra cost. The paper reports that this joint design outperforms forecasting-only and pseudo-anomaly baselines on five benchmarks, with larger AUC-PR gains than F1 gains, which it reads as evidence that the model produces a more reliable global ranking of anomaly likelihood rather than a better single operating point.","pith_inferences":["If the learned graph is the true carrier of the gains, then on higher-dimensional systems the fixed Top-$K$ sparsification ($K_g=5$, $K_{\\max}=3$) is likely to become a bottleneck, and an adaptive per-variable or per-dataset sparsity schedule is a natural testable extension.","The graphs are optimized for forecasting MSE, not for causal fidelity, so the NPE rankings should not be read as root causes without additional validation; an intervention or counterfactual test would be needed to separate structural evidence from spurious correlation.","Because the dual-view alerting head consumes only 13 fixed statistics per graph slice, edge-level detail is compressed away; a graph encoder that operates on the full adjacency matrix is a plausible upgrade if the structural signal plateaus.","A reasonable stress test of the paper's logic is to compare JAPE against a purely numerical forecaster on a benchmark where known anomalies are designed to have no numerical footprint but a clear dependency shift, since the framework's core prediction is that the structural view alone should still alert."],"forward_implications":["Structural precursors become usable evidence: alerts can fire on dependency rearrangements before numerical deviations become visible, addressing delayed detection.","Point-wise anomaly discrimination improves in the supervised-label setting, with reported average gains of 19.7% in F1 and 41.3% in AUC-PR over the strongest baselines, and the larger AUC-PR gains indicate a better-ranked anomaly score space.","Variable-level explanations come without extra training or inference: NPE reuses the predicted graphs, adds under 0.5% runtime, and improves MRR by 26.6% on the annotated datasets.","The two-stage design means the dependency graphs are learned purely from forecasting objectives and then consumed by both alerting and explanation, so the quality of the graph bounds the quality of both downstream outputs.","Lag range is consequential: sensitivity analysis shows that increasing the lag window from 1 to 3 patches helps, while too long a lag range degrades performance, making lag aggregation a tunable and load-bearing design choice."],"supporting_citations":[{"why":"Supplies the future-context forecasting baseline (FCM) that JAPE contrasts with its structural-view approach.","marker":"[45]"},{"why":"Supplies the self-supervised pseudo-anomaly baseline A2P and its supervised variant A2P-Sup used for comparison.","marker":"[28]"},{"why":"Supplies the channel-independent patched-transformer design that the temporal axis of DSTR follows and that also serves as a forecasting baseline.","marker":"[27]"},{"why":"Supplies iTransformer as a forecasting backbone and baseline for the prediction experiments.","marker":"[23]"},{"why":"Supplies the SMD dataset with anomaly labels and event-associated variable annotations used for prediction and explanation evaluation.","marker":"[31]"},{"why":"Supplies the WADI dataset, whose 123-variable setting exposes the high-dimensional risk of graph sparsification.","marker":"[2]"},{"why":"Supplies the MSL spacecraft-telemetry benchmark used in the prediction experiments.","marker":"[14]"},{"why":"Supplies the PSM server-monitoring benchmark used in the prediction experiments.","marker":"[1]"},{"why":"Supplies the EXATHLON benchmark, which provides the strongest single-dataset gains for the structural view.","marker":"[15]"},{"why":"Supplies the focal loss objective used to train the point-wise alerting head on imbalanced anomaly labels.","marker":"[22]"}],"fun_headline_variants":["Anomaly prediction rethinks: model dependency shifts","Weak signals? Track changing variable links","JAPE: anomaly early warning via dependency graphs","Structural deviations in time series predict anomalies","No extra model for explanations: reuse dependency graphs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the learned top-$K$ lag-aware directed graphs capture genuine dependency changes that precede anomalies, and if the contrast score mostly locks onto spurious correlations—especially as the number of variables grows—the structural view adds noise instead of signal, as the paper's own WADI result hints.","fun_headline_variants_meta":{"raw":{"variants":["Anomaly prediction rethinks: model dependency shifts","Weak signals? Track changing variable links","JAPE: anomaly early warning via dependency graphs","Structural deviations in time series predict anomalies","No extra model for explanations: reuse dependency graphs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000262,"raw_usage":{"total_tokens":1634,"prompt_tokens":1017,"completion_tokens":617,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":633,"completion_tokens_details":{"reasoning_tokens":548}},"tokens_in":633,"tokens_out":617,"duration_ms":7464,"temperature":1.0,"reasoning_tokens":548,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:26:47.990064+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a permutation control on SMD, MSL, and PSM where, at inference, the learned source variables in each graph slice are randomly rewired before the alerting head and the Graph Deviation Score are applied; if F1, AUC-PR, and MRR do not clearly drop relative to the true graphs, then the dependency structure is not carrying the reported gains.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the self-supervised pseudo-anomaly baseline A2P and its supervised variant A2P-Sup used for comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the WADI dataset, whose 123-variable setting exposes the high-dimensional risk of graph sparsification."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the MSL spacecraft-telemetry benchmark used in the prediction experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the PSM server-monitoring benchmark used in the prediction experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the EXATHLON benchmark, which provides the strongest single-dataset gains for the structural view."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the focal loss objective used to train the point-wise alerting head on imbalanced anomaly labels."}],"review_version":1}