acceptodds
Withdrawn from ICLR 2027

Breaking the Model Selection Barrier? Three Learning Paradigms for Supervised Ensembling in Time Series Anomaly Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Time series anomaly detection (TSAD) is a long-standing and extensively studied problem with applications across a large panel of domains. Despite the maturity of the field, recent benchmark studies have revealed that no single detection method consistently outperforms others across diverse datasets. While model selection approaches (i.e., choosing the best detector for a given scenario) have shown promising results, their effectiveness remains inherently limited by the performance ceiling of existing individual detectors. To address this limitation, supervised ensembling offers a promising path to surpass individual detectors by learning to combine their strengths. In this work, we unify and formalize the problem of supervised ensemble-based anomaly detection in time series, and introduce three principled strategies for learning such ensembles: (1) classical Machine Learning, (2) Reinforcement Learning, and (3) Genetic Programming. We perform a rigorous comparative evaluation across these strategies using identical model components, inputs, and experimental conditions to ensure fairness. Our findings not only highlight the strengths and trade-offs of each approach, but also identify Reinforcement Learning as a paradigm that can break the model selection barrier. Our experimental evaluation analyzes the robustness of this approach to distribution shift, label budget, and detector pool, pointing to promising directions for future research on this topic.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series, Ensembling, Anomaly Detection

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_X5MSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper learns, for each time series, a weight vector over 12 fixed anomaly detectors, and scores the series with the weighted sum of the detector scores. On TSB-UAD it compares three ways of learning the weights: gradient descent on an MSE proxy (ML), PPO with a VUS-PR reward (RL), and a genetic algorithm (GP). In-distribution, RL on catch22 features reaches a mean VUS-PR of 0.440, above the best model selector (0.382) and below the per-series oracle (0.489). ML and GP fall below the model selector.

Critical strengths
  1. Clear, well-posed question; the per-series Oracle is a useful ceiling for model selection.
  2. One pipeline (data, splits, detectors, network) for all three learning methods.
  3. Both in-distribution and leave-one-dataset-out settings are tested, and failures are reported.
  4. The per-series analysis of where the ensemble beats the Oracle (Fig. 5) is informative.
Critical weaknesses
  1. The main conclusion does not follow from the design. The paper credits the RL win to its "direct optimization of the true TSAD metric", but RL also differs from ML and GP in loss, training budget and tuning. Nothing isolates the learning paradigm as the cause.

  2. The budgets are unmatched. RL sees 720,000 states, while GP runs 20 generations of 20 solutions on 256 series each (about 100K series evaluations) over a much larger search space. Training takes about 25 h for RL and 5 h for GP (Fig. 6a). A fair comparison should match the number of metric evaluations or the wall-clock time.

  3. The ML proxy is weak. MSE against rare 0/1 labels mostly rewards low scores on normal points, while VUS-PR measures ranking. Ranking surrogates for AUC-PR exist (e.g. [1], [2]), as do pair-wise losses for weighting anomaly detectors [10], and they should be tried. Otherwise "ML" simply means "MSE".

  4. A baseline with a ranking (pair-wise) score function should be used for the classical ML method. Work on supervised ensembles of anomaly detectors [10] shows that learning detector weights from labels improves performance, and that a pair-wise comparison loss (an anomaly should score above a nominal) improves accuracy at the top of the ranking, which is also what the anomaly detection metric measures. The ML baseline here instead uses MSE against the 0/1 labels. A pair-wise or other ranking loss on the same weighted-sum model should be added, so that the comparison with RL and GP does not rest on a weak loss.

  5. The RL here is a one-step contextual bandit ("each state-action-reward tuple is treated independently"); PPO reduces to a score-function gradient with a learned baseline. Calling it a separate "paradigm" overstates the difference. Policy gradients for non-differentiable metrics are standard (e.g. self-critical training [8]), as are evolution strategies [9]; the RL and GP choices should be placed in this context.

  6. Simple baselines are missing: (a) one global weight vector learned on the training set, (b) score-level stacking on the 12 detector scores (as in XGBOD [3]), (c) average of the top-k detectors chosen by the model selector. The paper rules out stacking explicitly, but the title's claim needs at least (a) and (b). Related work on supervised and selective outlier ensembles (XGBOD [3], LSCP [4], SELECT [5], MetaOD [6]) is missing, as is the review of model selection for outlier detection [7].

  7. Variance is understated. The five seeds seem to vary only the training, not the 70/30 split. Figs. 7a and 7c use a single seed. It is also not said whether the Best MS value was re-run with the same VUS-PR code or copied from the earlier paper.

  8. Several details are missing: the number of series per split, the score normalization of the 12 detectors, what happens when all clipped weights are zero, and how ML/GP outputs are kept in [0, 1] (Sec. 4 says [0, 1], Appendix A.9 says [-1, 1]).

References

[1] Qi, Q., Luo, Y., Xu, Z., Ji, S. and Yang, T. (2021). Stochastic Optimization of Areas Under Precision-Recall Curves with Provable Convergence. NeurIPS 2021.

[2] Brown, A., Xie, W., Kalogeiton, V. and Zisserman, A. (2020). Smooth-AP: Smoothing the Path Towards Large-Scale Image Retrieval. ECCV 2020.

[3] Zhao, Y. and Hryniewicki, M. K. (2018). XGBOD: Improving Supervised Outlier Detection with Unsupervised Representation Learning. IJCNN 2018.

[4] Zhao, Y., Nasrullah, Z., Hryniewicki, M. K. and Li, Z. (2019). LSCP: Locally Selective Combination in Parallel Outlier Ensembles. SDM 2019.

[5] Rayana, S. and Akoglu, L. (2016). Less is More: Building Selective Anomaly Ensembles. ACM TKDD, 2016.

[6] Zhao, Y., Rossi, R. A. and Akoglu, L. (2021). Automatic Unsupervised Outlier Model Selection (MetaOD). NeurIPS 2021.

[7] Ma, Q. et al. (2023). The Need for Unsupervised Outlier Model Selection: A Review and Evaluation of Internal Evaluation Strategies. SIGKDD Explorations, 2023.

[8] Rennie, S. J., Marcheret, E., Mroueh, Y., Ross, J. and Goel, V. (2017). Self-critical Sequence Training for Image Captioning. CVPR 2017.

[9] Salimans, T., Ho, J., Chen, X., Sidor, S. and Sutskever, I. (2017). Evolution Strategies as a Scalable Alternative to Reinforcement Learning. arXiv:1703.03864.

[10] Das, S., Wong, W.-K., Dietterich, T., Fern, A. and Emmott, A. (2020). Discovering Anomalies by Incorporating Feedback from an Expert. ACM Transactions on Knowledge Discovery from Data (TKDD), 14(4), 1-32.

Questions for authors
  1. Does the RL advantage remain if ML uses a ranking surrogate loss for AUC-PR, such as a pair-wise loss [10]?

  2. How do the three methods compare under matched budgets?

  3. Where does D_val come from? Do the seeds resample the 70/30 split?

Reviewer_VpKrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper is mainly about ensembling different anomaly detectors rather than choosing only one to maximize the performance (i.e. VUS-PR in actual experiment), via three different methods including classic ML, RL, and genetic programming. Classic ML uses MSE as loss, which is not exactly the target metrics but an acceptable compromise to let the loss to be continuous and differentiable and can be optimized by gradient descent; RL depict the VUS-PR accuracy as reward and use gaussian sampling to get action W to do classic PPO gradient ascent; GP directly searches for the best M_theta that can get the best performance defined by the same metric. Three methods were compared under a same pipeline. Experiments found that RL on features can give best performance, sometimes even beyond oracle (the best detector selected after the detection). The main contribution of this paper is evaluating and comparing multiple baselines as well as deriving empirical conclusions.

Critical strengths

1.Clear visualizations, experimental results and analyses give baseline references and empirical clues for follow-up studies. Experiments are conducted over 16 datasets and 12 detectors, including in-distribution / out-of-distribution evaluations and analyses of label budget and detector pool to support the main claim that ensembling can go beyond simply selecting the best detector and RL within feature space goes beyond classic ML and GP implementations.

Critical weaknesses
  1. Although Appendix A.6 visualizes weight embeddings and Appendix A.11 reports detector-weight distributions, these analyses do not establish why RL performs better or how its detector preferences change with the pool or label budget. The reported ordering remains primarily a benchmark result, and a reader cannot tell when it would hold on a new dataset.
  2. Each paradigm has one implementation, so the comparison cannot carry a claim about paradigms. The RL arm is one PPO variant with Gaussian sampling and the GP arm is one genetic-programming setup, so what the numbers show is that this PPO implementation beats this GP implementation and this MSE-based learner. The title and the framing still rest on a comparison of three paradigms.
Questions for authors
  1. Were differences in scale among the detectors accounted for to ensure that the normalized weights have a fair effect? If yes, what norm is used to normalize the weights?
  2. Algorithm 1 says to take the first l points of a trajectory, whereas Appendix A.5 describes a segmentation method dividing the trajectory into windows of length l. Is there a contradiction?
  3. What does the weight clipping in Appendix A5 refer to? It is unclear whether the sampled actions or something else are clipped, and whether clipping occurs before or after normalization?
Reviewer_xFqhOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper formalizes supervised ensembling for time series anomaly detection and compares classical machine learning, reinforcement learning, and genetic optimization. The results demonstrate that learned ensembles can outperform practical model selectors and exceed the best individual detector on some time series.

Critical strengths

The unified framework provides a useful basis for studying supervised detector weighting. The evaluation covers distribution shifts, label budgets, detector removal, and computational efficiency. The analysis distinguishes practical model selection from the per-series oracle.

Critical weaknesses
  1. The three strategies primarily apply established optimization techniques without introducing a new learning mechanism.
  2. The claim of providing strong baselines is insufficiently supported. The detector pool and optimization methods largely consist of longstanding approaches, and comparisons with recent relevant methods are limited.
  3. The comparison confounds learning paradigms with optimization objectives. ML uses MSE while RL directly optimizes VUS-PR, so the results do not establish that reinforcement learning itself explains the improvement.
  4. The necessity of PPO is unclear because state–action–reward tuples are independent and no long-term decision process is involved.
  5. The claim of breaking the model selection barrier requires qualification. RL exceeds the oracle on some series, but its average VUS-PR remains below the oracle.
Questions for authors
  1. What justifies describing these approaches as strong baselines given the limited coverage of recent detectors and supervised ensembling methods?
  2. Does RL retain its advantage when competing methods use comparable objectives and computational budgets?
  3. Why is PPO preferable to simpler policy gradient methods, contextual bandits, or direct black-box optimization?
  4. Do the conclusions remain valid with a stronger and more recent detector pool?
  5. What methodological contribution distinguishes this work from applying established optimization algorithms to detector weighting?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.