Efficient cheap all-layer aggregation network for time-sensitive time series classification

초록

Time series classification is a key research area in the data mining community because it serves as a cornerstone for realizing various real-world applications. Existing deep learning approaches often rely on the final high-level features represented at the end of sequential layers, leading to information loss for complex and multiscale temporal dynamics. Recent studies adopted a layer-aggregation strategy to integrate multilevel temporal information. However, it is challenging for the final classifier to utilize temporal features from all the previous layers due to the substantial computational cost, making it challenging to develop time-sensitive tasks. To address this issue, we propose a cheap all-layer aggregation structure designed to enhance classification performance while maintaining computational efficiency. By leveraging learnable channel-wise transformation matrices and a scalable layer aggregation pool, the proposed model integrates information from all layers without increasing the computational costs. We first evaluated the proposed model against baseline methods in seven time-series classification tasks using six public benchmark datasets. To further emphasize our claim, we evaluated the proposed model on an additional six human activity recognition datasets. The proposed model achieved state-of-the-art performance while preserving computational efficiency. The source codes of the proposed network are publicly available at https://github.com/jgpark92/CALANet.

키워드