Deep time-series clustering via evolutionary learning and graph-based manifold learning

Deep time series clustering (DTC) methods have recently gained attention, but they often suffer from imbalanced clusters, sensitivity to initialization, and local optima due to their reliance on the KL-divergence-based loss. To overcome this, we propose a novel Deep Evolutionary Time Series Clustering (DETC) method, which uses an evolutionary search process, generating diverse candidate solutions in each iteration. These solutions are evaluated using a fitness function based on internal validation metrics in both raw time series and latent spaces. This enhances robustness and avoids sub-optimal solutions, leading to more stable and accurate clustering. DETC employs autoencoders for latent representation, but without proper constraints, models may learn poor representations, resulting in local optima, slow convergence, instability, and sensitivity to noise. To learn discriminative representations, DETC introduces a graph-regularized loss that preserves the topological structure of time series in both latent and reconstructed spaces. We conducted extensive experiments on 15 diverse time series datasets, including varying sample sizes, cluster counts, and sequence lengths for a comprehensive assessment. Experimental results demonstrate that DETC significantly outperforms existing state-of-the-art DTC benchmarks, showing its superior clustering performance and robustness. Among 11 compared methods, DETC obtains an average rank of 1.47 and leads to an average improvement of 11% in NMI compared to the best benchmark model. The code and data are available at: https://anonymous.4open.science/r/Deep-Evolutionary-Time-Series-Clustering-DETC-DD2D/README.md.