Abstract
HgAD is a self-adapting anomaly detection framework that jointly learns discrete hypergraph structures and models temporal trends and spatial relations among interdependent sensors using a hierarchical encoder-decoder architecture. Current approaches neglect higher-order dependencies within networks of interconnected sensors in multivariate time series data.
HgAD exploits relational inductive biases in hypergraph-structured data for one-step-ahead forecasting via self-supervised autoregressive task, detects anomalies based on forecast error, and provides root cause analysis through anomaly information propagation on computational hypergraphs along with offline optimal predictive control policy for remediation recommendations.
The framework introduces novel attention-based hypergraph convolution, local hypergraph pooling/unpooling, and hierarchical attention-based encoder-decoder techniques.
Key Highlights
Architecture
HgAD jointly learns discrete hypergraph structures capturing higher-order sensor dependencies and models temporal trends and spatial relations using a hierarchical encoder-decoder architecture. The framework leverages novel attention-based hypergraph convolution with local hypergraph pooling/unpooling to perform self-supervised one-step-ahead forecasting for anomaly detection, root cause diagnosis, and predictive control.
Key Contributions
- Hypergraph Structure Learning: Jointly learns discrete hypergraph structures capturing higher-order sensor dependencies.
- Hierarchical Architecture: Novel attention-based hypergraph convolution with local pooling/unpooling.
- Root Cause Analysis: Anomaly information propagation on computational hypergraphs for diagnosis.
- Predictive Control: Offline optimal control policy for remediation recommendations.
Results
HgAD demonstrates strong performance in anomaly detection and point estimation across multivariate time series benchmarks, accurately capturing temporal trends and spatial relations among interdependent sensors.
