AdaFSNet: Time Series Classification Based on Convolutional Network with a Adaptive and Effective Kernel Size Configuration

TL;DR

Time series classification often struggles to choose the right receptive field (RF) across varying sequence lengths. AdaFSNet addresses this by integrating an Omni-Scale OS-Block with multiple prime kernel sizes and a TargetDrop attention-based dropout to adaptively cover RFs, complemented by two dense blocks to exploit the selected kernels. The approach yields state-of-the-art accuracy on the UCR and UEA benchmarks, with fast convergence and robustness for both 1D and multivariate time series, while maintaining an end-to-end training regime. Overall, AdaFSNet provides an effective, scalable solution for automatic multi-scale feature extraction in time series classification without heavy preprocessing or extensive hyperparameter tuning.

Abstract

Time series classification is one of the most critical and challenging problems in data mining, existing widely in various fields and holding significant research importance. Despite extensive research and notable achievements with successful real-world applications, addressing the challenge of capturing the appropriate receptive field (RF) size from one-dimensional or multi-dimensional time series of varying lengths remains a persistent issue, which greatly impacts performance and varies considerably across different datasets. In this paper, we propose an Adaptive and Effective Full-Scope Convolutional Neural Network (AdaFSNet) to enhance the accuracy of time series classification. This network includes two Dense Blocks. Particularly, it can dynamically choose a range of kernel sizes that effectively encompass the optimal RF size for various datasets by incorporating multiple prime numbers corresponding to the time series length. We also design a TargetDrop block, which can reduce redundancy while extracting a more effective RF. To assess the effectiveness of the AdaFSNet network, comprehensive experiments were conducted using the UCR and UEA datasets, which include one-dimensional and multi-dimensional time series data, respectively. Our model surpassed baseline models in terms of classification accuracy, underscoring the AdaFSNet network's efficiency and effectiveness in handling time series classification tasks.

Figures (4)

• Figure 1: Analysis To RF Size. We trained the model on the Fiftywords dataset in UCR, where the horizontal coordinate represents the RF size, and we identified the kernel configuration of each 1D-CNN. For example, 9 (3-3-3-3) indicates that 1D-CNN has four layers and the receptive field size is 9, from the initial to the final layer, the kernel sizes are consistently set at 3 for each layer. As the receptive field size increases, there is a corresponding increase in accuracy. However, once a specific receptive field size is reached, further increasing it does not lead to improved accuracy; instead, it significantly escalates the computational demands of the model.
• Figure 2: AdaFSNet Pipeline. Goldbach's conjecture states that any even number can consist of two prime numbers. Base on this theory, with the OS-Block structure in the upper left of this image, which ensures the encompassment of all sizes of receptive fields. But not all prime numbers are suitable as kernel sizes. Therefore, we add a targeted dropout layer of attention module in the process to screen out more important RF, reduce redundancy and improve the accuracy of the model. In order to maximize the potential of PS-Block to cover all RF, We also add ResNet-like with 2Dense-Block module, whose kernel size is extracted from Dropout module. AdaFSNet achieves a series of SOTA performances on UCR & UEA datasets and we list the details in Table II.
• Figure 3: Influence of the number of Dense-Blocks
• Figure 4: Influence of the Dropout Module