This study presents a new ensemble approach for feature selection in time series forecasting, particularly in non-stationary and limited sample scenarios. The approach utilizes a hierarchical structure to exploit feature co-dependency. Initially, a machine learning model is trained using a subset of features. The model’s output is then updated using another algorithm with the remaining features to minimize target loss. This method overcomes the limitations of traditional feature selection methods and improves performance, scalability, and stability. The approach’s effectiveness is demonstrated on both synthetic and real-life datasets.
Publication date: 26 Oct 2023
Project Page: https://arxiv.org/abs/2310.17544v1
Paper: https://arxiv.org/pdf/2310.17544
