VGbel: An exploration of ensemble learning incorporating non-Euclidean structural representation for time series classification

Time series classification is an essential part of time series analysis research and has attracted generous researchers' attention. Representation learning and feature space expansion are vital issues in time series classification. A good representation can reveal the hidden information of the time series, while features could improve the performance of the classifier. Most proposed algorithms are based on the Euclidean structural representation of the time series. Inspired by the complex network analysis, it is realized that important structural features can be highlighted and characterized by non-Euclidean structural representation. Due to the lack of extensive evaluation on applying non-Euclidean structural representation and complex network theory to time series classification, an ensemble learning framework based on visibility graph representation (VGbel) is proposed to provide further exploration. Firstly, a directed series-to-graph transformer is constructed based on the visibility graph algorithm to map the time series into a non-Euclidean space. Secondly, the directed graph representation is subjected to multiscale feature extraction and is fused with the statistical features of the original time series. Furthermore, a hybrid enhanced ensemble model based on stacking is proposed for more stable performance. The proposed method has been extensively evaluated on UCR time series archive, and the experimental results demonstrate good classification accuracy and competitiveness.

Automated summary

Time series classification is important in time series analysis research and has gained attention from researchers. Representation learning and feature space expansion play vital roles in improving the performance of classifiers. Inspired by complex network analysis, non-Euclidean structural representation is introduced to highlight and characterize important features. To explore this approach further, an ensemble learning framework based on visibility graph representation (VGbel) is proposed, which incorporates directed series-to-graph transformation, multiscale feature extraction, and stacking-based ensemble modeling. Extensive evaluation on UCR time series archive demonstrates the effectiveness and competitiveness of the proposed method.