Applying graph neural network to multivariate time series anomaly detection

Abstract

The proliferation of data collection methods and technologies has underscored the importance and potential of data across various domains. Time series data, characterized by high dimensions and large volumes, serves as a valuable source for pattern discovery and information extraction in diverse fields. Anomaly detection algorithms for time series data have garnered significant interest due to their potential to serve as real-time monitors, aiding in incident tracking, outlier identification, and forecasting improvement. Motivated by the need to explore advanced anomaly detection techniques, this study investigates the performance of graph neural network-based anomaly detection models on multivariate time series data. Through comprehensive analysis of experiment results, it is evident that joint optimization and feature vector embedded graph attention mechanisms yield improved experimental outcomes. Notably, the combination of the two demonstrates enhanced capacity and sensibility in outputting meaningful error scores for unseen data. Additionally, evaluation method comparisons reveal the superiority of the epsilon search method in achieving higher F1 scores and lower latencies compared to the POT method. In conclusion, this project underscores the potential of graph neural network-based anomaly detection models in addressing real-world challenges associated with time series data analysis. By leveraging advanced techniques such as joint optimization and feature vector embedding, these models offer promising avenues for enhancing anomaly detection capabilities and improving real-time monitoring systems.