An Effective Dynamic Cost-Sensitive Weighting Based Anomaly Multi-classification Model for Imbalanced Multivariate Time Series

Sibo Qi, Juan Chen, Peng Chen, Jie Li, Wenyu Shan, Peian Wen

Abstract: Addressing imbalanced multivariate time series classification remains challenging due to skewed class distribution, resulting in suboptimal minority class classification. High dimensionality and temporal dependencies further complicate the task. We propose a novel model with dynamic cost-sensitive weighting to handle this. Our model employs multi-head self-attention and a transformer structure to capture dependencies. The proposed dynamic cost-sensitive weighting function enhances imbalanced multivariate time series handling with anomalies across classes. We comprehensively evaluated our model using KPI-monitored multivariate time series data via a microservice benchmark, comparing against baselines. Results underscore our model’s efficacy, especially in cloud computing and deep learning contexts.