Time2Graph: Dual Embedding and Nested-Graph Transformation for Performance Enhancement in Time Series Classification

초록

In this paper, we propose Time2Graph, a novel graph-based classification model designed to improve time series classification(TSC) performance in graph environments. Time2Graph applies a nested graph construction to effectively analyze time series in the graph domain. The macro graph, constructed from the entire dataset, represents global similarity and structural relationships among time series, while the micro graph captures local variations within individual time series. By jointly modeling inter-series global structures and intra-series local patterns, our nested graph design enhances embedding fidelity and maximizes classification performance. SimTSC is a recent approach that transforms time series into a macro graph and applies it to TSC. It adopts global sequence embedding(GSE), which focuses on the global information of each time series during feature embedding. However, GSE is less effective at preserving embedding fidelity in many datasets where local patterns are essential for capturing discriminative features, potentially degrading classification performance. To solve this problem, we propose localized graph embedding(LGE), which captures local information in node embeddings. LGE divides each time series into multiple subsequences, constructs a micro graph, and uses Graph2Vec to generate embedding vectors, effectively capturing local patterns and temporal variations. We further present multi-scale embedding, a weighted combination of GSE and LGE designed to maximize embedding fidelity. Through this dual embedding strategy, Time2Graph captures both fine-grained local information and global trends, leading to superior classification accuracy across diverse datasets. Experimental results on the UCR datasets demonstrate that Time2Graph outperforms SimTSC on approximately 84% of the datasets and achieves up to nearly twice the accuracy of SimTSC on some datasets.

키워드