SAFE-IML: Sparsity-aware feature extraction for interpretable machine learning with two-stage neural network modelling

In recent years, model interpretability has attracted significantly increasing attention and research interests from different backgrounds and perspectives. This paper focuses on interpretation of machine learning models, aiming to propose a new sparsity-aware feature extraction (SAFE) approach to significantly improve the interpretability of neural network models. The SAFE method includes the following two steps: 1) the first step starts with a set of features used for training machine learning models, to generate a significantly large number of new features; 2) with the awareness that augmented feature space is usually redundant, the second step is focused on dimensionality reduction to identify the most important features. These important features will then be used to train neural network models, enabling much better interpretability of learning results, as well as models themselves. The proposed method is referred to as Sparsity-Aware Feature Extraction for Interpretable Machine Learning (SAFE-IML). Two illustrative examples are provided to demonstrate the applicability and efficacy of SAFE-IML.