Adaptive feature fusion for time series classification

Abstract

Time series classification is one of the most critical and challenging problems in data mining, which exists widely in various fields and has essential research significance. However, to improve the accuracy of time series classification is still a challenging task. In this paper, we propose an Adaptive Feature Fusion Network (AFFNet) to enhance the accuracy of time series classification. The network can adaptively fuse multi-scale temporal features and distance features of time series for classification. Specifically, the main work of this paper includes three aspects: firstly, we propose a multi-scale dynamic convolutional network to extract multi-scale temporal features of time series. Thus, it retains the high efficiency of dynamic convolution and can extract multi-scale data features. Secondly, we present a distance prototype network to extract the distance features of time series. This network obtains the distance features by calculating the distance between the prototype and embedding. Finally, we construct an adaptive feature fusion module to effectively fuse multi-scale temporal and distance features, solving the problem that two features with different semantics cannot be effectively fused. Experimental results on a large number of UCR datasets indicate that our AFFNet achieves higher accuracies than state-of-the-art models on most datasets, as well as on the WISDM, HAR and Opportunity datasets, demonstrating its effectiveness.