ALIME: Autoencoder Based Approach for Local Interpretability

Abstract

Machine learning and especially deep learning have garnered tremendous popularity in recent years due to their increased performance over other methods. The availability of large amount of data has aided in the progress of deep learning. Nevertheless, deep learning models are opaque and often seen as black boxes. Thus, there is an inherent need to make the models interpretable, especially so in the medical domain. In this work, we propose a locally interpretable method, which is inspired by one of the recent tools that has gained a lot of interest, called local interpretable model-agnostic explanations (LIME). LIME generates single instance level explanation by artificially generating a dataset around the instance (by randomly sampling and using perturbations) and then training a local linear interpretable model. One of the major issues in LIME is the instability in the generated explanation, which is caused due to the randomly generated dataset. Another issue in these kind of local interpretable models is the local fidelity. We propose novel modifications to LIME by employing an autoencoder, which serves as a better weighting function for the local model. We perform extensive comparisons with different datasets and show that our proposed method results in both improved stability, as well as local fidelity.