Full Model Selection Problem and Pipelines for Time-Series Databases: Contrasting Population-Based and Single-point Search Metaheuristics

The increasing production of temporal data, especially time series, has motivated valuable knowledge to understand phenomena or for decision-making. As the availability of algorithms to process data increases, the problem of choosing the most suitable one becomes more prevalent. This problem is known as the Full Model Selection (FMS), which consists of finding an appropriate set of methods and hyperparameter optimization to perform a set of structured tasks as a pipeline. Multiple approaches (based on metaheuristics) have been proposed to address this problem, in which automated pipelines are built for multitasking without much dependence on user knowledge. Most of these approaches propose pipelines to process non-temporal data. Motivated by this, this paper proposes an architecture for finding optimized pipelines for time-series tasks. A micro-differential evolution algorithm (µ-DE, population-based metaheuristic) with different variants and continuous encoding is compared against a local search (LS, single-point search) with binary and mixed encoding. Multiple experiments are carried out to analyze the performance of each approach in ten time-series databases. The final results suggest that the µ-DE approach with rand/1/bin variant is useful to find competitive pipelines without sacrificing performance, whereas a local search with binary encoding achieves the lowest misclassification error rates but has the highest computational cost during the training stage.