- 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
- 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.
- 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.
- How were ℓ*, window size, m, k and α chosen without access to test labels? And what ℓ* was used for SMAP?
- Which stream and adjacency produce the test score? Could a classifier trained on the adjacency matrices alone separate the two training classes?
- Can you report mean ± std over several seeds, and explain the below-chance baseline AUCs on SWaT?
- Which timestep is used for the single-step reconstruction? Are the median and IQR in Eq. 13 computed over the test set?