[Paper Review] Rethinking 1D-CNN for Time Series Classification: A Stronger Baseline.

Wensi Tang, Guodong Long+4 more|arXiv (Cornell University)|Feb 24, 2020

Time Series Analysis and Forecasting26 references95 citations

TL;DR

This paper proposes Omni-Scale 1D-CNN (OS-CNN), a novel architecture that dynamically learns optimal kernel sizes for time series classification without grid search. By strategically configuring a small set of kernel sizes to cover diverse receptive fields, OS-CNN achieves superior performance across 85 UCR datasets, outperforming existing 1D-CNN baselines in accuracy, wins, and statistical significance.

ABSTRACT

For time series classification task using 1D-CNN, the selection of kernel size is critically important to ensure the model can capture the right scale salient signal from a long time-series. Most of the existing work on 1D-CNN treats the kernel size as a hyper-parameter and tries to find the proper kernel size through a grid search which is time-consuming and is inefficient. This paper theoretically analyses how kernel size impacts the performance of 1D-CNN. Considering the importance of kernel size, we propose a novel Omni-Scale 1D-CNN (OS-CNN) architecture to capture the proper kernel size during the model learning period. A specific design for kernel size configuration is developed which enables us to assemble very few kernel-size options to represent more receptive fields. The proposed OS-CNN method is evaluated using the UCR archive with 85 datasets. The experiment results demonstrate that our method is a stronger baseline in multiple performance indicators, including the critical difference diagram, counts of wins, and average accuracy. We also published the experimental source codes at GitHub (this https URL).

Motivation & Objective

• To address the inefficiency and suboptimal performance of grid search for kernel size selection in 1D-CNNs for time series classification.
• To theoretically analyze the impact of kernel size on 1D-CNN performance in capturing salient temporal patterns.
• To develop a novel architecture that enables 1D-CNN to learn the optimal receptive field scale during training, reducing reliance on manual hyperparameter tuning.
• To establish a stronger baseline for 1D-CNN in time series classification by improving model capacity through intelligent kernel size configuration.

Proposed method

• Propose a novel Omni-Scale 1D-CNN (OS-CNN) architecture that integrates multiple kernel sizes in a single, unified convolutional block.
• Design a specific kernel size configuration strategy using a minimal set of carefully selected kernel sizes to cover a wide range of receptive fields.
• Integrate the multi-scale kernels into a single convolutional layer to enable end-to-end learning of the most relevant scale during training.
• Use a shared feature map across different kernel sizes to maintain computational efficiency while enhancing representational capacity.
• Train the model end-to-end with standard optimization to allow the network to automatically learn the most effective kernel size for each time series pattern.
• Evaluate the model on the full UCR Time Series Archive (85 datasets) using standard metrics including accuracy, critical difference diagrams, and win counts.

Experimental results

Research questions

• RQ1 How does kernel size selection impact the performance of 1D-CNN in time series classification tasks?
• RQ2 Can a 1D-CNN model learn the optimal kernel size dynamically during training without relying on grid search?
• RQ3 What is the minimal set of kernel sizes needed to effectively cover diverse receptive fields in time series data?
• RQ4 How does the proposed OS-CNN architecture compare to standard 1D-CNN baselines in terms of accuracy and robustness across diverse time series datasets?
• RQ5 Does the proposed method serve as a stronger baseline than existing 1D-CNN approaches in time series classification?

Key findings

• OS-CNN achieves higher average accuracy than existing 1D-CNN baselines across the 85 UCR datasets.
• The proposed method records more wins than baseline models in pairwise comparisons across the UCR benchmark.
• OS-CNN demonstrates superior performance in critical difference diagrams, indicating statistically significant advantages over standard 1D-CNNs.
• The model's performance is robust across diverse time series data, confirming the effectiveness of the designed kernel size configuration.
• The method reduces the need for time-consuming grid search by learning optimal kernel sizes end-to-end, improving training efficiency.
• The source code is publicly available on GitHub, enabling reproducibility and further benchmarking.