Adaptive interactive network component ensemble for streaming data with concept drift

Husheng Guo, Zhijie Wu, Zhengqi Liu, Shuai Zhang, Wenjian Wang

Abstract: Concept drift is an important issue in streaming data mining. At present, most concept drift processing methods face some challenges such as slow model convergence after drift occurring and weak representation ability for data stream with complex distributions. In this paper, an adaptive interactive network component ensemble method is proposed to solve the problem of streaming data with concept drift. First, the convolution features with different scales are concatenated, and the channel relationship is modelled to construct concatenated features by applying the attention mechanism. Then the shallow input features and concatenated features are fused adaptively by the gating mechanism. The multi-scale feature extraction module, channel attention mechanism and gating mechanism are used to construct a multi-scale feature fusion network component, the features of diversity and discriminability can be obtained from this component. Second, the mutual learning is used to learn from representative components with different properties, which enhances representation ability of shallow components and convergence ability of deep components. The experimental results on simulation data streams and visual object tracking data streams show that the proposed method can effectively deal with concept drift of streaming data and quickly converge to the new distribution when drift occurs. It outperforms traditional adaptive methods in dealing with concept drift and can improve the representation ability of data with complex distribution and real-time generalization performance of the model.

External IDs:dblp:journals/datamine/GuoWLZW25