{"id":"6aad1345-10dd-456b-ab83-713e42f39f7e","arxiv_id":"2411.16427","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A GAN-plus-reinforcement-learning detector learns to remove suspicious events from continuous-time sequences without clean data, beating the PPOD baseline on synthetic and real benchmarks.","lead":"This paper proposes an unsupervised method for spotting abnormal events in continuous-time event sequences: a reinforcement-learning agent removes suspected outliers and a discriminator judges whether the edited sequence looks real. It claims to be the first approach that works without clean training data and reports higher accuracy than the PPOD semi-supervised baseline on four datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'unsupervised' claim is not yet supported: Appendix G tunes hyperparameters on training AUROC that uses ground-truth labels, so the reported advantage over PPOD may be a selection artifact.","rationale":"The Reader's weakest assumption concerns the additive-only anomaly model: the generator can only remove points, so missing events cannot be represented or detected, despite the Related Work claiming to detect 'abnormal occurrences and absence of events.' I agree this scope mismatch exists and should be fixed by narrowing the claims. However, the Reader's own strongest_claim explicitly restricts attention to 'injected additive outliers under the tested conditions,' so the additive-only concern does not directly refute the central claim as stated. The label-based hyperparameter tuning in Appendix G is more load-bearing because it threatens both the 'unsupervised' and 'more accurate' parts of the central claim. If the authors selected hyperparameters by training-set AUROC, then ground-truth labels influenced model selection, contradicting the paper's claim of running 'without any supervision from humans.' This is not merely a rhetorical point: Figure 8 shows that AUROC varies substantially with learning rates, so the chosen configuration matters. A concrete re-evaluation with label-free model selection could settle whether the reported advantage over PPOD is a property of the algorithm or an artifact of supervised tuning. The paper has internal consistency in its experiments and the narrow additive-outlier comparison to PPOD appears plausible, so I do not recommend rejection; the appropriate action remains conditional acceptance pending this check. My disagreement with the Reader is therefore partial: we both flag the unsupported breadth of the claims, but I consider the label-tuning issue to be the single most consequential threat to the central claim.","tokens_in":11195,"tokens_out":7331,"duration_ms":76155,"concrete_test":"Re-run the Table 1 experiments with hyperparameters selected by a purely label-free criterion (e.g., held-out discriminator reward, with no ground-truth labels used at any point) or with fixed defaults from Appendix G, and compare test AUROC to PPOD after applying the same selection procedure. If GAN-RL does not beat PPOD on a majority of datasets under this protocol, the reported gains are attributable to supervised model selection rather than to the unsupervised algorithm.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that GAN-RL performs unsupervised outlier detection and outperforms PPOD. Appendix G states that 'For all the experiments, the hyper-parameters were tuned on final AUROC scores on the training set.' Since AUROC is computed from ground-truth outlier labels, and Section 3.1 says those labels are 'not accessible by any of the methods during training,' this model-selection step is a supervised procedure. The reported results therefore do not demonstrate that a fully unsupervised pipeline would choose these hyperparameters, nor that the method would beat PPOD in a genuinely label-free setting. Figure 8 shows that performance is sensitive to learning rates, so the hyperparameter choice is consequential. Without a label-free selection rule or evidence that default hyperparameters are adequate, the 'unsupervised' and 'more accurate' components of the central claim are not established. The additive-only anomaly model is a limitation, but the strongest claim can be narrowed to injected additive outliers; the label-tuning issue attacks the central claim as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GAN-RL, an unsupervised method for detecting outlier events in continuous-time event sequences. A generator, trained with PPO, sequentially decides for each event whether to remove it, producing a 'corrected' sequence; a discriminator tries to distinguish corrected sequences from real (possibly corrupted) sequences, and its real-class probability is used as the reward for the generator. After training, the generator's removal probabilities are used as an online per-event outlier score. The method is evaluated on two synthetic datasets (Poisson and Hawkes) and two real-world datasets (MIMIC, Taxi), with outliers injected as additional Poisson points. Reported test AUROC values (Table 1) show GAN-RL outperforming PPOD, RND, and LEN, and the appendix provides ablations on corruption level, attention, the GAN reward, and end-to-end training.","tokens_in":11335,"tokens_out":5138,"duration_ms":45570,"significance":"If the central empirical claim holds, the paper would be the first to demonstrate unsupervised, online detection of individual event outliers in continuous-time event streams, a practically useful extension of the semi-supervised PPOD method. The paper has clear strengths: evaluation on four datasets with multiple seeds, ablations that isolate the contributions of attention and the adversarial reward, a causal attention mechanism that preserves online applicability, and a comparison against a relevant state-of-the-art baseline. However, the unsupervised claim is not yet fully established because hyperparameters are selected using ground-truth labels, and the anomaly model covers only additive point outliers rather than the broader class of anomalies mentioned in the text. These issues are fixable but currently limit what the experiments demonstrate.","major_comments":[{"comment":"The statement that 'For all the experiments, the hyper-parameters were tuned on final AUROC scores on the training set' is in tension with the unsupervised claim. AUROC is computed with ground-truth outlier labels, which Section 3.1 explicitly says are not accessible during training. This makes model selection a supervised procedure, so the advantage over PPOD in Table 1 may be a selection artifact rather than a property of unsupervised training. The issue is consequential: Figure 8 shows strong sensitivity to learning rates. Please report results with a label-free selection rule (e.g., fixed defaults, or model selection using only unsupervised criteria), or explicitly reposition the contribution as unsupervised training with supervised hyperparameter selection.","section":"Appendix G"},{"comment":"The anomaly model covers only additive point outliers. The generator's action space is binary keep/remove, and all experiments inject outliers by generating extra Poisson points that are merged with clean sequences. The abstract and Related Work claim that the method detects 'abnormal occurrences and absence of events', but the proposed action space cannot represent missing events or altered timestamps; it can only delete existing points. Please either narrow the claim to additive point outliers or extend the action space and evaluation to include missing or shifted events.","section":"Section 2.3 and Appendix C.2"},{"comment":"PPOD is the main baseline in Table 1, but the paper does not describe how PPOD was trained for these experiments, which hyperparameters were used, or whether its hyperparameters were tuned under the same protocol (label-based or label-free) as GAN-RL. Without this information, the statement that GAN-RL 'can detect event outliers more accurately than the state-of-the-art approaches' is not fully supported. Please report the PPOD training protocol and ensure the baseline is tuned under comparable conditions.","section":"Section 3 (baseline setup)"}],"minor_comments":[{"comment":"The sentence about learning-rate sensitivity refers to 'Fig 8 (a) & (b)', but the generator learning rate is shown in panel (c); also 'discrimiantor' is misspelled.","section":"Appendix F"},{"comment":"Reference [23] contains a stray space in 'V olodymyr'; please correct it.","section":"References"},{"comment":"The heading 'Comparison of asymptotic training performance' is vague; specify the number of episodes used and whether the numbers correspond to the same random seeds as Table 1.","section":"Table 2"},{"comment":"The row 'Maximum Time Length=10' in Table 3 needs units and a statement of how longer sequences are handled (truncation, splitting, or padding).","section":"Appendix G"}],"recommendation":"major_revision","confidential_remarks":"The paper is a workshop-style manuscript whose main contribution is empirical. The label-based hyperparameter tuning in Appendix G is the key obstacle: it undermines the 'unsupervised' wording of the central claim. If the authors rerun the comparison with a label-free selection rule or demonstrate that default hyperparameters give comparable results, the paper would be much stronger. I would also ask the authors to verify the 'first' claim against the broader anomaly-detection literature before final submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a plausible first crack at fully unsupervised point-level outlier detection for event sequences, and the GAN+RL combination works on the datasets tested. But the 'unsupervised' claim is partly undermined by label-based hyperparameter tuning in Appendix G, and the method only handles additive outliers even though the text claims to address absence of events. With reframing and a cleaner model-selection story, it could become a solid paper.\n\nWhat's actually new: the problem formulation. PPOD needs clean training data, Shchur et al. and Zhu et al. work at sequence level, and I've not seen a fully unsupervised point-level setting in event sequences before. The method itself is a reasonable combination of known pieces — cLSTM, causal attention, PPO, and a GAN-style discriminator — and the key intuition (generator mistakes create useful augmented anomalies for the discriminator) is sensible. The ablations are well-chosen: attention matters, a learned reward beats a Wasserstein distance reward, and the method degrades gracefully as corruption increases. The reported AUROC improvements over PPOD are consistent across four datasets.\n\nThe soft spots are real. First, the unsupervised framing is generous: Appendix G says hyperparameters were tuned on training-set AUROC, and AUROC uses the ground-truth labels. That is a supervised model-selection step, and Figure 8 shows the method is sensitive to learning rates, so the advantage over PPOD could be partly a selection artifact. It's not fatal — most ML papers tune on labeled validation sets — but it should be acknowledged, and the paper should either show a label-free selection rule or demonstrate robustness across a grid of default hyperparameters. Second, the generator can only remove points. The Related Work says the method detects 'abnormal occurrences and absence of events,' but a missing event has no point to remove. So the 'absence' claim is unsupported. The paper should either narrow the scope to additive outliers or extend the action space to insert points. Third, the baseline set is thin: PPOD is the only serious comparison, and its training setup here isn't described. RND and LEN are trivial. No code is provided, which is a reproducibility problem.\n\nWho it's for: people working on anomaly detection in temporal point processes or event streams. It's a workshop-level contribution with a good core idea, but it needs major revision before it supports the claims as stated. I'd send it to a serious referee; they'd likely request the same fixes.","headline":"A clever first attempt at fully unsupervised point-level outlier detection, but the label-tuned hyperparameters and additive-only action space mean the 'unsupervised' and 'absence of events' claims are ahead of the evidence.","tokens_in":11925,"tokens_out":5020,"would_cite":false,"duration_ms":47165,"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":"The paper claims an unsupervised GAN-RL loop detects event outliers more accurately than the semi-supervised PPOD baseline, with no clean training data.","keywords":["unsupervised outlier detection","event sequences","temporal point processes","generative adversarial networks","reinforcement learning","continuous time","online anomaly detection","PPO"],"falsifier":"Train GAN-RL on sequences whose anomalies are missing events rather than added points — delete a fraction of normal events from otherwise clean histories and ask the detector to flag those positions — and check whether the AUROC stays above chance; if it does not, the claim that the method detects both abnormal occurrences and absences of events is refuted.","tokens_in":10941,"feed_emoji":"⏱️","tokens_out":11512,"duration_ms":97576,"temperature":0.7,"pith_summary":"This paper proposes what it describes as the first fully unsupervised method for detecting outlier events in continuous-time sequences, where the training data are unlabeled and may be corrupted. The framework pairs a reinforcement-learning generator that decides, point by point, whether to delete an event with a GAN-style discriminator that tries to tell the corrected sequence from the real one. The learning signal is an asymmetry: correct deletions make the corrected sequence look like the clean majority and hard to distinguish, while wrong deletions of normal points are easy for the discriminator to catch, so the generator's mistakes become training data for the discriminator. After training, the generator works as an online point-level outlier detector on unseen sequences. The paper reports higher AUROC than the semi-supervised baseline PPOD on synthetic Poisson and Hawkes data and on MIMIC ICU admissions and NYC taxi trip data with injected additive outliers.","feed_headline":"GAN-RL beats semi-supervised baseline at event outlier detection","feed_subtitle":"Generator-discriminator loop turns correction mistakes into a training signal for online, label-free detection.","key_machinery":"The load-bearing machinery is the adversarial correction loop. A continuous-time LSTM encoder with a causally masked attention layer and layer normalization turns each event history into states; a PPO-based RL agent reads those states and samples a keep/remove action per point, so its policy is a per-point outlier probability; a spectrally normalized discriminator scores the whole corrected sequence against real sequences. The identity that carries the argument is the correction contrast: if the generator removes true outliers, the corrected sequence is statistically indistinguishable from the clean majority, but if it removes a normal point, the discriminator can detect the artifact, and that contrast supplies the reward signal for training the otherwise non-differentiable deletion policy.","core_discovery":"The central discovery, stated on the paper's own terms, is that an adversarial loop can be turned from a sequence-generation tool into an unsupervised outlier detector: the RL generator's deletion policy directly estimates the probability that each event point is an outlier, and the discriminator's ability to separate corrected sequences from real, mostly clean sequences acts as the reward that shapes that policy. The authors position this as the first unsupervised approach for abnormal-event detection in continuous time and report that it outperforms the state-of-the-art semi-supervised baseline PPOD, which assumes clean training data. On 100 test sequences unseen during training, GAN-RL achieves AUROC 0.631 versus 0.55 on Poisson, 0.610 versus 0.512 on Hawkes, 0.778 versus 0.583 on MIMIC, and 0.647 versus 0.548 on Taxi under the paper's additive-outlier injection protocol.","pith_inferences":["A reader should not generalize the reported accuracy to anomaly types outside the remove-only action space: missing events, substituted events, and shifted timestamps are not representable by this generator, so new experiments would be needed before claiming detection of 'absence of events'.","The same correction-contrast principle could be extended to richer editing actions, such as inserting missing points or nudging timestamps, which would turn the framework into a general unsupervised data-repair model.","The learned deletion probability could double as a per-point anomaly score and, aggregated, as a sequence-level score, linking this work to sequence-level anomaly detection benchmarks.","Because the paper evaluates only injected additive outliers with ground-truth labels, a natural stress test is deployment on naturally occurring anomalies in real event logs, where the outlier-generating mechanism is unknown."],"forward_implications":["Event-forecasting pipelines could train directly on polluted sequences and use the learned deletion policy as an automated cleaning step, removing the need for a separate outlier-free corpus.","Because all attention is causally masked, the trained generator can be deployed online on streaming data, flagging each new event before the next one arrives.","The method replaces the semi-supervised assumption of clean training data with a weaker majority-clean assumption; the ablation suggests performance is preserved down to roughly 60% clean sequences.","The discriminator supplies a learned, data-dependent reward, so the detector does not require a hand-specified intensity model of normal behavior as likelihood-based baselines do.","At the other extreme, when no clean sequences remain ($\\beta = 0$), the learned AUROC falls to about 0.5, confirming that the method's signal is the contrast between corrected and majority-clean real data."],"supporting_citations":[{"why":"Defines the semi-supervised event outlier detection task and supplies the PPOD baseline that the paper compares against.","marker":"[15]"},{"why":"Provides the clipped PPO algorithm used to train the generator's point-removal policy.","marker":"[27]"},{"why":"Supplies the continuous-time LSTM (cLSTM) encoder that turns event histories into states for the generator and discriminator.","marker":"[19]"},{"why":"Supplies the adversarial generator-discriminator training concept on which the framework is built.","marker":"[7]"},{"why":"Spectral normalization applied to discriminator layers to stabilize adversarial training.","marker":"[22]"},{"why":"Attention mechanism used with a causal mask in the encoder to keep the detector online.","marker":"[4]"},{"why":"Layer normalization applied after the attention layer in the generator encoder.","marker":"[2]"}],"fun_headline_variants":["First unsupervised outlier detection for continuous-time event sequences","GAN-RL turns correction mistakes into a training signal for anomaly detection","Unsupervised GAN-RL outperforms semi-supervised baseline on event outliers","Adversarial loop detects event outliers without labels in continuous time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every anomaly is an extra event point that should be removed, and that clean sequences are the majority of the data; anything else — missing events, altered timestamps, substituted events — is outside what the generator can represent or detect.","fun_headline_variants_meta":{"raw":{"variants":["First unsupervised outlier detection for continuous-time event sequences","GAN-RL turns correction mistakes into a training signal for anomaly detection","Unsupervised GAN-RL outperforms semi-supervised baseline on event outliers","Adversarial loop detects event outliers without labels in continuous time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000941,"raw_usage":{"total_tokens":4008,"prompt_tokens":919,"completion_tokens":3089,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":535,"completion_tokens_details":{"reasoning_tokens":3025}},"tokens_in":535,"tokens_out":3089,"duration_ms":19966,"temperature":1.0,"reasoning_tokens":3025,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:07:30.823578+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train GAN-RL on sequences whose anomalies are missing events rather than added points — delete a fraction of normal events from otherwise clean histories and ask the detector to flag those positions — and check whether the AUROC stays above chance; if it does not, the claim that the method detects both abnormal occurrences and absences of events is refuted.","supporting_citations":[{"cited_title":"Event outlier detection in continuous time","cited_arxiv_id":null,"evidence_quote":"Defines the semi-supervised event outlier detection task and supplies the PPOD baseline that the paper compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the continuous-time LSTM (cLSTM) encoder that turns event histories into states for the generator and discriminator."},{"cited_title":"Generative adversarial nets","cited_arxiv_id":null,"evidence_quote":"Supplies the adversarial generator-discriminator training concept on which the framework is built."},{"cited_title":"Neural Machine Translation by Jointly Learning to Align and Translate, May 2016","cited_arxiv_id":null,"evidence_quote":"Attention mechanism used with a causal mask in the encoder to keep the detector online."}],"review_version":1}