acceptodds
Withdrawn from ICLR 2027

DiffTSAD: Diffusion-Guided Dynamic Graph Learning for Unsupervised Multivariate Time-Series Anomaly Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multivariate time-series anomaly detection (MTSAD) is critical for monitoring complex systems, while it suffers from three key limitations: the anomaly reconstruction paradox under severe class imbalance, epoch-dependent optimization conflicts where evolving reconstruction fidelity inverts supervision signals during co-training, and rigid static graph topologies that fail to capture anomaly-induced structural shifts. We propose DiffTSAD, a framework that fundamentally decouples generative anomaly simulation from discriminative learning. By leveraging a frozen pre-trained diffusion model as a stationary perturbation prior, DiffTSAD injects severity-calibrated perturbations to establish a distribution-invariant binary classification paradigm, eliminating the non-stationary supervision of epoch-dependent pseudo-labeling methods. Diffusion-induced residuals further guide dynamic graph construction to adaptively reweight inter-variable dependencies, while a topology-aware Graph Attention Network with chunk-wise dilated temporal convolutions enables fine-grained spatio-temporal representation learning. Experiments on three real-world benchmarks demonstrate that DiffTSAD achieves state-of-the-art or competitive performance.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series, Dynamic graph learning, Anomaly detection

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_5DRFOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper tackles unsupervised multivariate time-series anomaly detection by turning it into a proxy classification problem. A diffusion model is pre-trained on the normal training data and then frozen. During training, each normal window is paired with a noised copy obtained from the diffusion forward process at a fixed level ℓ*, and the network learns to tell the two apart (clean = 0, noised = 1). Each stream gets its own graph: a sigmoid of z-scored XᵀX, plus, for the noised stream, a boost to the rows and columns of the variables that changed most under noise. Both streams then go through a GAT variant with learnable node embeddings, dilated temporal convolutions over chunks, and an MLP head. At test time the score is a 50/50 mix of the classifier output and a single-step diffusion reconstruction error. The authors’ main argument is that because the generator is frozen, the supervision doesn’t drift across epochs the way it does in co-training methods like SPAGD. A few propositions in the appendix back this up (a VC bound, a Weyl bound on the graph injection, a TV-distance decomposition).

Critical strengths
  • I appreciated that the authors drop point adjustment and report AUPRC. That’s the right call given how inflated numbers in this area usually are. They are also upfront in the appendix that the propositions are motivation, not guarantees. This makes the results easier to trust and to interpret, which matters for my assessment because it lets me see clearly where the method does and doesn’t work.
  • The efficiency comparison (Table 5) is a nice touch. It shows the extra pre-training stage doesn’t make the method more expensive than SPAGD overall. If the core method held up, this would be a real point in its favour.
Critical weaknesses
  • My main concern is that the pre-trained diffusion model doesn’t seem to be involved in creating the training data at all. Unless I’m misreading Eq. 1 and Algorithm 1, X̃ comes straight from the closed-form forward process, so ε_θ is never called. The complexity section even confirms no network evaluation is needed here. The training task is therefore just “clean vs. Gaussian-noised”, and pre-training or freezing the model makes no difference to it. Stationarity (Prop. 1) then holds for any fixed noise augmentation, so it’s not really a property of the proposed design. The noise is also quite heavy. I did a rough calculation with the stated linear schedule:
    • At ℓ = 0.3, ᾱ comes out around 0.4, i.e. noise with std ≈ 0.77 on [0,1]-scaled data.
    • At ℓ = 0.7, the level picked for SWaT, ᾱ is under 0.01, so the “perturbed” sample is basically pure noise.

It’s hard to see why learning to separate data from noise would transfer to real anomalies. The ablations seem to agree: the classifier alone gets 56.3 AUC on MSL, below COPOD, k-NN and OCSVM. I’m also worried the graph gives the label away. Since r_i ≥ 0, σ(r_i) is always at least 0.5, so the noised stream always carries large injected edges that the clean stream never has.

  • The results don’t convince me.
    • On SMAP, AUC is 51.4 and AUPRC is 15.1 with a 13.1% anomaly rate. That’s essentially random, and so are all the SMAP ablations.
    • On SWaT, COPOD wins on AUPRC and AT wins on F1, and the average gap over COPOD is about 0.3 points.
    • Improvements are quoted as relative percentages, which makes them look bigger than they are.
    • There are no repeated runs or standard deviations, so differences like 64.80 vs 64.55 can’t be read as meaningful.
    • Several baselines get AUCs well below 50 on SWaT (e.g. NCAD 19.0), which makes me suspect something went wrong in running them.
    • The biggest single drop in the ablation table comes from removing “smoothing”, about 10 AUC points on MSL. That component isn’t described, and as far as I can tell it isn’t applied to the baselines. This looks like the kind of post-processing the paper says it avoids.
    • ℓ* and window size appear to be chosen from curves computed on test labels.
    • Evaluating only on MSL/SMAP/SWaT is also limiting, given the known problems with these benchmarks, which the authors themselves cite.
Questions for authors
  1. Is X̃ generated purely by forward noising, as in Algorithm 1, or noised and then denoised by ε_θ, as Figure 2 suggests? If it’s the former, how does the method compare to the same setup without any pre-trained model? Only the reconstruction term at test time would change.
  2. What is the temporal smoothing, exactly? Was it applied to the baselines? I’d like to see Table 1 with smoothing either applied to all methods or to none.
  3. How were ℓ*, window size, m, k and α chosen without access to test labels? And what ℓ* was used for SMAP?
  4. Which stream and adjacency produce the test score? Could a classifier trained on the adjacency matrices alone separate the two training classes?
  5. Can you report mean ± std over several seeds, and explain the below-chance baseline AUCs on SWaT?
  6. Which timestep is used for the single-step reconstruction? Are the median and IQR in Eq. 13 computed over the test set?
Reviewer_6WDgOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

In this paper the authors present DiffTSAD, a diffusion-guided anomaly detection framework. The presented model designed to detect multivariate time-series anomalies consists of two parts, namely a frozen pre-trained diffusion model and a discriminative classifier. The training injects severity-calibrated perturbations to establish a distribution-invariant binary classification and eliminate epoch-dependency.

Critical strengths

S1: The proposed DiffTSAD method performs well, either better or competitive to 14 compared baselines.

S2: In comparison to SPAGD on the MSL dataset, DiffTSAD has an overall faster execution time. A time complexity analysis is included in the paper.

S3: The paper contains significant analysis and ablation study.

Critical weaknesses

W1: Because of the frozen pre-trained diffusion backbone, I would think that the model cannot adjust to slowly drifting normal sample distributions. Please see my question Q1 below.

W2: The frozen backbone is exclusively trained on normal samples. In a practical setting, how would a user obtain a time series that guaranteed does not contain any outliers?

W3: The paper does not describe the limitations of the proposed approach.

Questions for authors

Q1: Since the backbone is a frozen pre-trained diffusion model, how does the proposed method work when the distribution of the normal samples is slowly changing?

Q2: The frozen backbone is exclusively trained on normal samples. In a practical setting, how would a user obtain a time series that guaranteed does not contain any outliers?

Reviewer_SH1LOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DiffTSAD, an unsupervised multivariate time-series anomaly detection framework that combines diffusion-based perturbations with dynamic graph learning. The method generates synthetic anomalies using a fixed Gaussian diffusion process and constructs residual-guided adjacency matrices to capture inter-variable dependencies. A spatio-temporal encoder integrates graph attention and dilated temporal convolutions for binary classification.

Critical strengths
  1. The paper focuses on unsupervised multivariate time-series anomaly detection, a practically relevant task. It considers limitations of reconstruction-based methods and static graph structures, although whether the proposed mechanisms effectively overcome these limitations remains unclear.
Critical weaknesses
  1. Limited methodological novelty and substantial overlap with SPAGD. DiffTSAD largely follows SPAGD[1], including synthetic anomaly generation, residual-based node scoring, top-m% variable selection, adjacency enhancement, and dual-stream spatio-temporal classification. In particular, Eqs. (2) and (4) closely match SPAGD's Eqs. (5) and (6). The remaining changes mainly involve the perturbation source and attention design. It is unclear what substantial methodological contribution remains beyond modifications to an existing framework.

  2. Inconsistency in the most relevant baseline results. Table 1 reports SPAGD AUC scores of 53.63, 44.48, and 64.28 on MSL, SMAP, and SWaT, whereas the original SPAGD paper reports 66.50, 62.38, and 86.30. The originally reported SPAGD results outperform DiffTSAD on all three datasets. Both studies report the same dataset statistics and exclude point adjustment.

  3. The learned diffusion model is unnecessary for the proposed perturbation-based training. Despite presenting the frozen pre-trained diffusion model as a central contribution, Eq. (1) generates synthetic anomalies using only a fixed Gaussian noise schedule, without evaluating the learned diffusion model. Thus, diffusion pre-training is not needed for the claimed stationary proxy supervision. The learned model contributes only to the reconstruction score at inference, making the claimed diffusion-guided training mechanism misleading.

  4. Proposition 1 establishes only that fixed perturbations produce a stationary training distribution; it does not prove that co-training causes optimization conflicts or that freezing the generator improves convergence. Similarly, the residual-based graph enhancement does not directly identify anomaly-induced changes in variable dependencies. Removing the dynamic graph reduces AUC by only 0.25 and 0.32 percentage points on MSL and SMAP, respectively. These results do not substantiate the claimed importance of the proposed mechanisms.

  5. Despite the claimed robustness, DiffTSAD achieves only 51.44% AUC on SMAP, barely above random ranking, and its improvements on several other metrics are small. The paper also does not clearly specify how dataset-specific diffusion levels are selected or which data are used to compute the median and IQR in Eq. (13). These omissions raise concerns about the reliability and reproducibility of the reported results.

[1] Self-Perturbed Anomaly-Aware Graph Dynamics for Multivariate Time-Series Anomaly Detection, Cat et.al, NeurIPS 2025.

Questions for authors

Minor Comments:

The manuscript is generally readable, but several presentation issues should be addressed. Some statements are stronger than the supporting evidence, particularly the claims about eliminating optimization conflicts and the necessity of individual components. The interpretation of improved AUC and AUPRC as evidence of better probability calibration in Section 4.2 is also inaccurate. In addition, the descriptions of score fusion and temporal smoothing in Section 4.3 are not fully consistent with the formulation in Eq. (13).

The paper would also benefit from more careful proofreading. For example, the phrase "a topology-aware spatio-temporal representation learning" in Section 3 is grammatically incomplete, and the conclusion contains the incorrect construction "outperforms state-of-the-art baselines or highly competitive performance." In Appendix A.2, the final step of the proof of Proposition 2 appears to omit the +1 term, although it is retained in Eq. (15). The reference (line 518) to SPAGD should also be corrected from 2026 to 2025.

Questions

  1. Why are the SPAGD results in Table 1 substantially lower than those reported in the original paper, despite using the same datasets and point-wise evaluation protocol? Were the baselines reproduced under different settings?

  2. Since Eq. (1) generates perturbations using only a predefined Gaussian noise schedule, what does the pre-trained diffusion model contribute to discriminative training? Would the classifier achieve the same performance without diffusion pre-training?

  3. Given the substantial overlap, which specific design changes account for the improvements over SPAGD? Can their contributions be isolated experimentally?

  4. Is there direct evidence that epoch-dependent perturbations cause the claimed label inversion or training instability? How does the proposed decoupling resolve this issue beyond simply fixing the perturbation distribution?

  5. How are dataset-specific diffusion levels and other hyperparameters selected? Are anomaly labels used during validation? Also, which data are used to compute the median and IQR in Eq. (13)?

Reviewer_hye5OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

DiffTSAD combines diffusion-based perturbations, residual-guided dynamic graphs, and a spatio-temporal classifier for unsupervised multivariate time-series anomaly detection. Its final anomaly score combines classification probability with diffusion reconstruction error. Experiments cover MSL, SMAP, and SWaT.

Critical strengths
  1. The evaluation covers three real-world benchmarks and reports AUC, AUPRC, and point-wise F1 without Point Adjustment, alongside component ablations and efficiency comparisons.
Critical weaknesses
  1. The central contribution is inconsistent with the specified algorithm. Equation (1) and Algorithm 1 generate training perturbations exclusively through a fixed noise schedule and Gaussian noise, without using the pretrained network. Section 4.5 explicitly confirms that this operation requires no network evaluation. Consequently, epoch-invariant supervision follows from fixed noise injection, independently of diffusion pretraining or freezing. The learned denoiser contributes to reconstruction scoring at inference, but this does not substantiate the central claim that a frozen learned prior guides anomaly simulation and stabilizes classifier training.
  2. The experimental evidence does not establish the claimed importance of dynamic graph learning. In Table 2, removing the dynamic graph reduces AUC by only 0.25 and 0.32 percentage points on MSL and SMAP, respectively. By comparison, the “w/o Smoothing” variant reduces MSL AUC by 9.89 points. Section 4.3 states that this variant removes both IQR normalization and temporal smoothing, yet the temporal smoothing operation is unspecified in Equation (13) and Algorithm 1. A substantially more influential scoring component is therefore incompletely documented, while the assertion that the proposed dynamic graph is indispensable remains unsupported.
Questions for authors
  1. Where do the pretrained diffusion parameters enter training-time perturbation generation, and why is freezing necessary when Equation (1) contains no learned parameters?
  2. What temporal smoothing operation produces the reported results, and is the same operation applied to the baselines?

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

Related papers

Discussion (0)

Sign in to comment.