Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·May 25·5 min read
VACE: Learning Geometrically Structured Representations for Time Series Anomaly Detection
TL;DR
Detecting anomalies in multivariate time series (data streams with multiple variables) is difficult because abnormal patterns are rare and unlabeled. VACE (Velocity-Aligned Channel Embeddings) uses self-supervised learning to map normal behavior into a compact, geometrically coherent region in embedding space, then flags deviations using position and velocity scores.
✦ Why It Matters
Engineers can deploy simpler, more interpretable anomaly detection systems that explicitly control embedding geometry rather than relying on complex contrastive heuristics.
Key Takeaways
- VACE uses velocity-consistency training without negatives or synthetic anomalies, simplifying self-supervised learning.
- Combine positional (Mahalanobis) and directional (velocity-bank) scores multiplicatively for robust anomaly flagging.
- Channel-aware encoding processes each variable distinctly, improving representation quality for multivariate data.
- Explicit geometric structure control outperforms indirect contrastive pair-sampling on TSB-AD-M benchmark.
Full Summary
Anomaly detection in multivariate time series—identifying unusual patterns in data streams with multiple variables—is critical for applications like system monitoring where abnormal events are rare, labeled examples don't exist, and missing anomalies is costly. Existing self-supervised methods use contrastive learning (comparing similar and dissimilar examples) to embed time series patches into a latent space where normal data clusters together, but these approaches indirectly shape the geometric structure through heuristic pair sampling, offering no explicit control over how tightly normal representations group or whether distances are meaningful.
VACE addresses this by training a channel-aware encoder (one that processes each data variable distinctly) using a velocity-consistency objective—ensuring normal trajectories move smoothly and align in embedding space—without requiring negative examples or synthetic anomalies. At test time, VACE combines a Mahalanobis positional score (measuring statistical distance from the normal cluster) with a velocity-bank directional score (measuring dynamic atypicality) multiplicatively to flag anomalies.
Despite its simplicity, VACE outperformed more complex methods on the TSB-AD-M benchmark, suggesting that explicit geometric structure is more effective than indirect contrastive approaches.