Comparative Approaches to Time-Series Clustering

Clustering is an unsupervised machine learning technique in which data is grouped into segments such that observations within a segment are similar while those across segments differ. The technique has a wide range of applications including customer segmentation, anomaly detection, and feature engineering. While traditional methods applied to cross-sectional data employ static features and distance-based algorithms such as k-means or hierarchical clustering, time-series data is additionally complex due to its sequential structure. Aggregating transactional data into tabular form often neglects dynamic behavioural patterns, underscoring the need for methodologies that explicitly account for temporal dependencies. This research note reviews and compares different approaches to time-series clustering, including summary-statistic transformations, “native” methods such as Dynamic Time Warping (DTW), feature-based frameworks like RFM and Catch22, and deep learning models like Deep Temporal Clustering (DTC). Using a transactional dataset from Victory Farms, a sustainable aquaculture business in Kenya, we apply all methods to real-world customer data. The results highlight how time-series clustering yields richer insights than cross-sectional approaches, while also serving as a practical case study that combines theory, code, and application.