Interpretable machine learning for shared feature identification and time series prediction

Abstract

When building a predictive model with sub-datasets from diverse locations, scenarios, or participants, a single model may not capture the unique characteristics of each sub-dataset. However, creating individual models for each dataset can be time-consuming and may overlook shared features. In this article, a Common Structure Neural Network (CSNN) model is introduced to address these issues. The model includes a new feature selection layer that identifies critical shared factors influencing multiple outputs, allowing for a shared model structure and reduced training costs, while accurately representing the diversity within each sub-dataset. The effectiveness of the model is demonstrated through one simulation and two real-world case studies on air pollution and stock prices. The experiments show that the model improves prediction accuracy and efficiency compared to other methods. Additionally, it enhances interpretability by revealing correlations and interactions across different locations, offering valuable insights.