Evolutionary Feature Selection for Time-Series Forecasting

In machine learning, feature selection is crucial for pinpointing the key subset of features that enhances interpretability and preserves or boosts the model's original performance. Filter methods, which assess features using statistical metrics, are particularly notable. Recently, a novel metric called Conditional Dependence Coefficient has been proposed to measure the dependence between subsets of variables. This paper introduces a novel filter feature selection method that integrates the Conditional Dependence Coefficient metric with an evolutionary algorithm to find the optimal feature subset. This approach combines the adaptability of genetic algorithms with the strength of an intuitive metric. Unlike many filter-based methods, our technique does not rely on parameters directly linked to the number of features (like thresholds). Moreover, it evaluates the collective merit of feature subsets rather than individual significance. We conducted tests on six different multivariate time-series datasets to address the forecasting challenge, determining the relevant lags. Considering no selection as baseline, experimental results indicate that our approach is competitive in terms of efficacy while demonstrating a reduction in the number of features selected.