Causal interpretability and uncertainty estimation in mixture density networks*

Neural network implementations have predominantly been a black box lacking both in interpretability and estimation of uncertainty. In this study, we propose a novel causal attribution methodology for mixture density networks wherein we outline a framework to compute the causal effect of each feature on the target variable along with the associated uncertainty in the attribution. Our approach allows for the prediction and causal estimation tasks, along with the uncertainty estimation, to be integrated within the same architecture thus obviating the need to train a separate causal model. We report experimental results on two real-world problems comprising of studying the causal impact of bio markers on diabetes progression and the causal impact of certain key features on ecommerce sales. We also evaluate our approach on an open source simulated dataset and compare our results against multiple state-of-the-art baselines.