High-Dimensional Time Series Classification Based on Similarity Measure

초록

High-dimensional time series classification often requires complex models, extensive featureengineering, or deep learning architectures. We propose a simple and training-free classification method that leverages a recent similarity measure called the Dimension InsensitiveEuclidean Metric (DIEM), developed by Tessari and his colleague in 2024. Unlike traditionaldistance-based or learning-based approaches, our method directly compares input vectorswith labeled instances using DIEM, without requiring model training or parameter tuning. Toevaluate its effectiveness, we applied it to benchmark datasets with varying dimensionality. In particular, on the Olive Oil dataset, the method achieved an accuracy of 87.2% and wasover 2,000 times faster than a Gramian angular field (GAF)-based CNN. On the Meat dataset,it experienced only a 1.7% point reduction in accuracy compared to GAF, while still beingmore than 1,400 times faster. Our results show that the proposed method not only matches orexceeds the accuracy of complex models but also offers significant computational advantagesacross a variety of classification tasks.

키워드