Tree-based learning for multivariate time series clustering

Time series data, capturing sequential measurements over time, is of great significance to diverse domains such as finance, meteorology, and biology. While the importance of time series clustering is evident, particularly in understanding temporal patterns, existing methods face challenges in multivariate time series settings. The focus of this thesis is on predictive clustering trees, a promising approach, yet unexplored in multivariate time series. In contrast, neural networks are effective but lack the interpretability for real world applications. To address these gaps, a novel predictive clustering tree algorithm for time series, that can effectively handle the intricacies of multivariate time series is presented. The proposed approach incorporates a consideration of intra- and inter-cluster distances, ensuring both cohesion and separation. This hybrid strategy enhances the quality of clustering, enabling the capture of diverse patterns while maintaining cluster distinctiveness. A key strength is the algorithm's flexibility, demonstrated through effective handling of both univariate and multivariate time series. Interpretability is often neglected in clustering models but is a critical factor in real-world decision-making. The proposed algorithm not only produces clusters based on similarity but also provides insights, aligning with the need for transparency in applications. Experimental evaluations conducted on various datasets, including a real-world application in electrical engineering, showcase promising results in clustering quality and performance.