acceptodds
Under review as a conference paper at ICLR 2027

SPECGATE: CROSS-DOMAIN CONTRASTIVE PROTO TYPES WITH SPECTRAL GATING FOR UNSUPERVISED TIME-SERIES ANOMALY DETEC TION

Abstract

Unsupervised anomaly detection on multivariate time series is dominated by reconstruction- and density-based models, yet most treat the temporal and spec tral views in isolation or rely on a single contrastive pretext that can collapse in narrow industrial regimes. We propose SPECGATE, a model that brings three complementary signals to the same point in time. A shared patch transformer learns a time-domain view through masked reconstruction; a learned frequency band gate produces a channel-adaptive spectral view whose reconstruction error grows when energy appears outside the normal spectral support; an InfoNCE ob jective aligns the two views in the joint space; and a normalised prototype mem ory banks the normal manifold so that off-manifold anomalies incur large resid ual distance. The three evidence channels are fused with channel-wise standard ised scores. SPECGATE reaches competitive PA-F1 on six standard benchmarks (SMD, MSL, SMAP, PSM, SWaT, WADI) without any anomaly labels at train ing time and with < 2M parameters. Beyond empirical gains, we provide a short analysis showing that the gate acts as a data-driven band-selector and that the prototype memory yields a quantifiable separation between normal and abnormal embeddings under standard concentration assumptions.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multivariate time series, unsupervised anomaly detection, contrastive learning, spectral gating, prototype memory

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_kWqYOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SpecGate, a framework for unsupervised multivariate time-series anomaly detection that combines masked temporal reconstruction with gated spectral reconstruction, cross-view contrastive alignment, and prototype memory. It integrates reconstruction errors and prototype distances to generate point-level anomaly scores.

Critical strengths
  1. The method attempts to combine time–frequency representation learning with prototype-based anomaly detection.
Critical weaknesses
  1. Limited novelty and an apparent engineering combination, with insufficiently articulated technical contributions.
  2. The ablation studies are incomplete and do not adequately verify the necessity of each design choice.
  3. Baseline comparison is incomplete in the main evaluation. Consequently, the claimed competitive performance cannot be assessed.
  4. The writing and presentation quality still need improvement.
Questions for authors
  1. Could the authors provide complete benchmark results?
Reviewer_hcqSOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper is not a research paper in an evaluable format, but rather an unfinished draft that reads as if generated by automated writing tools. It cobbles together four standard components—masked block reconstruction, gated spectral autoencoder, InfoNCE loss term, and prototype memory module—yet reports no benchmark test results whatsoever.

Critical strengths

None. The submission contains no experimental evidence, no valid theoretical result, and no demonstrated technical contribution. There is therefore no substantive strength that could support acceptance.

Critical weaknesses

Lack of baseline test results. Table 2 reports the PA-F1 metric across six datasets, but the results for CATCH, KAN-AD, MRD, CrossAD, and SpecGate are all marked as “–”; Table 3 (the complete ablation study) similarly shows “–”. No specific numerical values for SpecGate appear anywhere in the manuscript.

Questions for authors
  1. Please provide complete reports of Table 2 and Table 3, including the actual numerical values for all six datasets, all baseline methods under the same protocol, and all ablation studies. If the data are not yet available, the paper should not be submitted.

  2. Since the model is channel-independent, how does it detect anomalies that manifest solely as violations of inter-channel correlations? Please include the results of ablation studies on the cross-channel mixing module.

Reviewer_qU17OpenReview ↗
Rating
1
Confidence
1
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes an approach to anomaly detection in time-series data. The method is based on a combination of time-domain and frequency-domain features, which are combined using a contrastive scheme. The method is unsupervised and experiments are performed on multiple time-series datasets with a variety of SOTA competing methods.

Critical strengths

The overall approach seems well thought out, with elements of novelty.

Critical weaknesses

The main weakness is the lack of experimental results. Table 2 contains no values for the SpecGate row, and Table 3 is empty.

In addition, there are some odd formatting decisions. Section 2 is very list-like, which is unusual for a Related Work section. Figure 1 is a block diagram, which is hard to read and fairly unattractive.

Most of the equations are unnumbered, and Algorithm 1 makes reference to equation numbers that don't exist in the text, or are incorrect.

The two propositions are not contextualized in the paper. They may indeed be correct and important to the work, but their relation to the main method needs to be explained in the body of the paper - they can't be expected to stand alone. If space is a concern, it would be more appropriate to simply explain the importance of these mathematical results in the body of the paper, and leave the formal propositions and proofs for the appendices.

The bullet list of data set in Sec. 4.1 is a waste of space, and provides very little information, which could more efficiently be added to Table 1.

Questions for authors

Complete Tables 2 and 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.