An efficient GAN-based predictive framework for multivariate time series anomaly prediction in cloud data centers

Sibo Qi, Juan Chen, Peng Chen, Peian Wen, Xianhua Niu, Lei Xu

Abstract: Recently, a growing amount of time series data has been collected in cloud data centers, making anomaly detection for multivariate time series analysis increasingly necessary. However, extracting meaningful features from multivariate time series remains challenging due to the limited amount of labeled data and highly complex temporal correlations. Additionally, many unsupervised deep learning methods often result in a high false alarm rate. This study proposes a new unsupervised multivariate time series anomaly prediction model called the Predictive Wasserstein Generative Adversarial Network with Gradient Penalty (PW-GAN-GP). Our model adopts both Wasserstein Distance and Gradient Penalty, making the adversarial training more stable and helping the generator’s output to more closely resemble the real data. Moreover, a novel anomaly score function combining reconstruction, discrimination, and prediction errors is used to improve precision while maintaining recall. The experimental results on four public cloud computing datasets demonstrate that our proposed PW-GAN-GP outperforms the suboptimal baseline, with improvements of 22.11% and 13.47% in precision and F1 scores, respectively.