Shapley-based feature augmentation

Improving the predictive performance of machine learning models is the desired goal in many tasks and domains. The predictive performance of the learning algorithm is directly affected by the input features it receives. Feature augmentation is aimed at enhancing the quality of models by adding informative features to the original data. Explainable AI methods are typically used to explain the results of machine learning models. Recently, these methods have also been used to improve models' predictive performance. In this study, we examine the benefit of incorporating the explanations obtained by an explainable AI method as augmented features. In particular, we propose SFA - Shapley-Based feature augmentation, a two-stage ensemble learning method that uses out-of-fold predictions and their corresponding Shapley values as augmented features for each instance. Shapley values, which are obtained without domain expertise, reflect the importance of the original features to each prediction and consider their interactions with all other features. Experimental results demonstrate the superiority of our proposed method, SFA, against several feature augmentation methods on multiple public datasets with various characteristics.