Dynamic Graph Regularization for Multi-Stream Concept Drift Self-Adaptation

Ming Zhou, Jie Lu, Pengqian Lu, Guangquan Zhang

Abstract: Concept drift is an inevitable problem in non-stationary stream environments, due to evolving data distributions. In practical applications, multi-stream is more complex than single-stream, yet they have received little attention. Addressing concept drift while mining correlations between streams poses a significant challenge. Research focuses on capturing correlations between streams using graph neural networks (GNNs), providing valuable insights. However, these methods fix the correlation graph structure after training, failing to adapt to the new data distributions with dynamic correlations during testing. To bridge this gap, we propose a novel framework named Multi-stream Self-adaptation based on Graph Regularization (MSGR). A new GNN architecture is proposed to capture deep spatio-temporal correlations and learn a correlation graph structure without pre-defined graphs. Each graph node represents a stream and the graph is constructed through Gumbel sampling and an adaptive matrix from stream pairs. Thus we attain a base high-performance GNN for multi-stream multi-step prediction. To adapt to the new data distribution during testing, we design a self-adaptation mechanism by assigning dynamic learning weights for newly arriving samples. Larger learning weights are assigned to relevant samples when drift occurs. The self-adaptation is accomplished by the sub-graph updating and the proposed graph regularization. Error-based drift detection is integrated, and when drift occurs, the weight for sub-graph updating increases by adjusting the regularization coefficient. Thus, MSGR maintains high self-adaptation performance and accurate prediction results consistently regardless of the type and degree of concept drift. Comprehensive testing on real-world and synthetic datasets shows that MSGR achieves state-of-the-art performance.

External IDs:dblp:journals/tkde/ZhouLLZ24