File size: 6,159 Bytes
d3e570a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
---
license: mit
language:
- en
pretty_name: VITS-AD Evaluation Suite
size_categories:
- 10K<n<100K
task_categories:
- time-series-forecasting
tags:
- anomaly-detection
- time-series
- evaluation
- benchmark
- frozen-vision
- regime-analysis
- negative-results
- neurips-2026
---

# VITS-AD Evaluation Suite

**Companion artifact for the NeurIPS 2026 Evaluations & Datasets (E&D) Track
submission *VITS-AD: A Regime-Aware Evaluation Suite for Frozen-Vision
Time-Series Anomaly Detection*.**

This dataset is **not a new corpus**. It bundles the *evaluation outputs*
produced by the VITS-AD pipeline and the raw-space Mahalanobis baseline so
that future work can:

1. **Reproduce paper tables and statistical tests** without re-running the
   full vision pipeline.
2. **Run paired comparisons** (Wilcoxon, paired-99 UCR) directly on the
   per-window scores.
3. **Audit the regime classification** (amplitude vs. structural) against the
   underlying evidence artifacts.

The submission is **double-blind**; this dataset card is anonymous. Source
code is at `https://github.com/evaldataset/VITS-AD` (reviewer-routed via
`anonymous.4open.science`).

## Contents

| Folder              | Purpose                                                   | Size  |
|---------------------|-----------------------------------------------------------|-------|
| `regime_labels/`    | Per-dataset regime annotation + classifier features       | <1 KB |
| `ledgers/`          | JSON ledgers underlying main-paper claims                 | 64 KB |
| `ucr_canonical/`    | UCR 109/99 aggregated and per-series metrics              | 168 KB |
| `multiseed_scores/` | Per-window scores + labels for 5 seeds × {LP,RP} × {PSM,MSL,SMAP}, no model weights | 9.8 MB |
| `sample_renderings/`| Pipeline diagram and regime-gain figures                  | 1 MB  |

Total: ~11 MB.

## What this is for (E&D Track scope)

The submission's contribution is **benchmark analysis and evaluation
methodology**, not a new dataset. We therefore distribute:

- The **regime axis** (amplitude vs. structural) along which vision
  rendering does and does not pay off.
- The **paired-99 UCR comparison** that demonstrates Wilcoxon $p<10^{-7}$
  in favour of the vision pipeline on the structural univariate regime.
- The **per-seed scores** that allow re-running paired Wilcoxon and
  bootstrap confidence intervals on the multiseed PSM/MSL/SMAP runs.
- The **calibration and FPS ledgers** that back the compute-disclosure and
  CalibGuard tables in the paper supplement.

We do **not** redistribute the raw benchmark datasets (SMD, PSM, MSL, SMAP,
UCR Anomaly Archive). License and download paths for those upstream
benchmarks are listed in the paper's Asset Credits table.

## Files

### `regime_labels/regime_labels.json`

Per-dataset regime label, channel count, raw vs. VITS-AD AUC-ROC, and the
five classifier features. The accompanying notes record the in-sample
classifier accuracy ($90.1\%$), the majority-class baseline ($88.1\%$), and
the leave-one-dataset-out CV collapse to chance — i.e., the regime axis is
**descriptive**, not a deployable predictor.

### `ledgers/`

| File | Backed claim |
|------|--------------|
| `improved_ensemble_results.json` | SMD 28-entity macro AUC-ROC for VITS-AD vs. raw Mahalanobis |
| `multiseed_results.json` | $n=5$ seed mean ± std for PSM/MSL/SMAP × {LP, RP} |
| `multiseed_ensemble_summary.json` | Rank-mean ensemble across renderers per dataset |
| `optimized_ensemble.json` | Oracle renderer-adaptive scoring |
| `calibguard_multidataset.json` | Realized FAR vs. target FAR (empirical diagnostic) |
| `fps_benchmark.json` | FPS, parameter count, and compute disclosure |
| `clip_backbone_comparison.json` | DINOv2 vs. CLIP backbone ablation |
| `ucr_results.json` | Legacy UCR aggregate (paired-99 in `ucr_canonical/`) |
| `view_disagree_sweep.json` | Cross-view disagreement scoring sweep |

### `ucr_canonical/`

Authoritative UCR ledgers used by every UCR claim in the paper:

| File | Description |
|------|-------------|
| `summary.json`     | 109-series VITS-AD aggregate |
| `paired_99.json`   | 99-series paired comparison vs. raw Mahalanobis |
| `combined_109.json`| Per-series VITS-AD scores |
| `per_series.json`  | Per-series metric breakdown |
| `eligible_list.json` | List of the 109 eligible UCR series |
| `ucr_canonical.json` | Combined manifest |

### `multiseed_scores/`

Layout: `multiseed_scores/{psm,msl,smap}/{line_plot,recurrence_plot}/seed_{42,123,456,789,2024}/`

Each leaf directory contains:

- `scores.npy` — per-window anomaly score (float64)
- `labels.npy` — per-window ground-truth label (int64, $\{0,1\}$)
- `metrics.json` — AUC-ROC, AUC-PR, best-F1, F1-PA for that seed

Model checkpoints (`best_model.pt`) are intentionally **not** redistributed
to keep the bundle compact; they can be regenerated from the source repo.

### `sample_renderings/`

PDFs of the pipeline diagram (Figure 1), per-dataset regime gain
(Figure 3a), and the regime-map scatter (Figure 4 in the supplement).

## Reproducing the paper's statistical tests

```python
import json, numpy as np
from scipy.stats import wilcoxon

base = "multiseed_scores/psm/line_plot"
seeds = [42, 123, 456, 789, 2024]
aucs = []
for s in seeds:
    m = json.load(open(f"{base}/seed_{s}/metrics.json"))
    aucs.append(m["auc_roc"])
print(f"PSM-LP mean ± std: {np.mean(aucs):.4f} ± {np.std(aucs, ddof=1):.4f}")

# Paired-99 UCR Wilcoxon (vision vs. raw Mahalanobis)
paired = json.load(open("ucr_canonical/paired_99.json"))
stat, p = wilcoxon(paired["vits_ad_auc"], paired["raw_maha_auc"])
print(f"Paired-99 UCR Wilcoxon p={p:.2e}")
```

## License

MIT. All redistributed JSON ledgers, regime annotations, and rendered
example PDFs are original work of the (anonymous) authors and are released
under MIT alongside the source repository.

## Citation

```bibtex
@inproceedings{vitsad2026,
  title  = {{VITS-AD}: A Regime-Aware Evaluation Suite for Frozen-Vision Time-Series Anomaly Detection},
  author = {Anonymous},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track},
  year   = {2026}
}
```