A Multiple Objective Particle Swarm Optimization Algorithm for Time Series Segmentation

Time series segmentation is aimed at representing a time series by using a set of segments. In most cases, time series segmentation can reduce approximation error by increasing number of segments, implying a conflict between these two factors. Different techniques have been applied to time series segmentation and achieved good results, yet the simultaneous optimization of the number of segments and approximation error is seldom considered. For this, this paper proposes a Multiple Objective Particle Swarm Optimization algorithm for Time Series Segmentation (MOPSOTTS). MOPSOTTS improves a Dynamic Exploitation Barebones Particle Swarm Optimization (DBBePSO) algorithm, aiming for multiple objective optimization in time series segmentation. To generate segmentation solutions with varying numbers of segments and different cut points, MOPSOTTS designs crossover and mutation operators to update the particle vectors throughout the iterative process. On two time series from real-world problems and four time series from UCR repository, we compare MOPSOTTS with three traditional algorithms for time series segmentation, conduct a controlled study on MOPSOTTS, and analyze the computation time overhead of these algorithms. According to the experimental results, MOPSOTTS generally outperforms three traditional algorithms for time series segmentation, and the improved DBBePSO algorithm and the crossover and mutation operators can effectively enhance the performance of MOPSOTTS.