# HFN: Heterogeneous Feature Network for Multivariate Time Series Anomaly Detection

**Authors:** Jun Zhan, Chengkun Wu, Canqun Yang, Qiucheng Miao, Xiandong Ma  
**arXiv:** [2211.00277](https://arxiv.org/abs/2211.00277)  
**Date:** 2022-11-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2211.00277v2/

## Key points

- The potential for network or physical attacks on industrial equipment or computer systems to cause massive losses highlights the need for quick and accurate anomaly detection based on monitoring data, especially multivariate time-series (MTS) data.
- Most existing approaches do not consider the heterogeneity between variables, which may have different and distinctive edge distributions depending on their type (continuous numerical variables, discrete categorical variables or hybrid variables).
- A novel semi-supervised anomaly detection framework called HFN has been proposed in this paper to address this issue.
- The HFN is based on a heterogeneous feature network that learns heterogeneous structure information from a mass of unlabeled time-series data to improve the accuracy of anomaly detection.
- It uses attention coefficients to provide an explanation for detected anomalies.
- The proposed approach combines the embedding similarity subgraph generated by sensor embedding and feature value similarity subgraph generated by sensor values to construct a time-series heterogeneous graph that fully utilizes the rich heterogeneous mutual information among variables.
- Experiments were conducted using four sensor datasets from real-world applications, demonstrating that the proposed approach detects anomalies more accurately than baseline approaches.
- This provides a basis for rapid positioning of anomalies and can help prevent massive losses caused by network or physical attacks on industrial equipment or computer systems.
- Overall, this paper presents an important contribution towards improving anomaly detection in MTS data through consideration of variable heterogeneity using HFNs with attention coefficients.
- The results demonstrate that HFN outperforms existing methods in terms of accuracy and explainability for detecting anomalies in MTS data sets from real world applications.

## Summary

The potential for network or physical attacks on industrial equipment or computer systems to cause massive losses highlights the need for quick and accurate anomaly detection based on monitoring data, especially multivariate time-series (MTS) data. However, most existing approaches do not consider the heterogeneity between variables, which may have different and distinctive edge distributions depending on their type (continuous numerical variables, discrete categorical variables or hybrid variables). To address this issue, a novel semi-supervised anomaly detection framework called HFN has been proposed in this paper. The HFN is based on a heterogeneous feature network that learns heterogeneous structure information from a mass of unlabeled time-series data to improve the accuracy of anomaly detection. Additionally, it uses attention coefficients to provide an explanation for detected anomalies. The proposed approach combines the embedding similarity subgraph generated by sensor embedding and feature value similarity subgraph generated by sensor values to construct a time-series heterogeneous graph that fully utilizes the rich heterogeneous mutual information among variables. A prediction model containing nodes and channel attentions is jointly optimized to obtain better time-series representations. This approach fuses state-of-the-art technologies of heterogeneous graph structure learning (HGSL) and representation learning. Experiments were conducted using four sensor datasets from real-world applications, demonstrating that the proposed approach detects anomalies more accurately than baseline approaches. This provides a basis for rapid positioning of anomalies and can help prevent massive losses caused by network or physical attacks on industrial equipment or computer systems. Overall, this paper presents an important contribution towards improving anomaly detection in MTS data through consideration of variable heterogeneity using HFNs with attention coefficients. The results demonstrate that HFN outperforms existing methods in terms of accuracy and explainability for detecting anomalies in MTS data sets from real world applications.