Optimized symbolic aggregate approximation methods for lightweight and interpretable multivariate time series classification

Multivariate time series classification often requires balancing predictive performance with model simplicity and transparency, especially in domains constrained by interpretability or computational resources. This paper introduces a lightweight and interpretable symbolic framework for multivariate time series classification built upon the Symbolic Aggregate approXimation (SAX) representation. Despite SAX’s well-established merits in terms of dimensionality reduction, noise robustness, and interpretability, its practical effectiveness is highly sensitive to the choice of transformation parameters which are typically set heuristically. To address this limitation, we leverage the DIRECT global optimization algorithm to provide a principled, deterministic tuning framework that adapts these parameters to each dataset without resorting to ensemble or deep architectures. We propose four feature extraction strategies that vary in parameter sharing and channel aggregation, including a Multichannel Intelligent Icon (MII) frequency vector that captures cross-dimensional dependencies. Experiments on 24 benchmark datasets from the UEA archive demonstrate accuracy competitive to strong baselines with significant runtime advantages over computationally demanding alternatives, validated through Friedman and Nemenyi post-hoc statistical analysis. The proposed framework is particularly well-suited for applications where efficiency, trustworthiness, and explainability are essential, demonstrating that globally optimized symbolic representations constitute a viable alternative to more complex models for transparent multivariate time series classification.