MANE: Model-Agnostic Non-linear Explanations for Deep Learning Model

Deep learning methods have shown great potentiality in the credit card fraud detection field. Unfortunately, they cannot explain their predictions, while a financial company expects to know the reason that a transaction is decided to be fraud. There have been some studies of the interpretability of deep learning methods in other fields, but we find that their effects are not very good when we directly use them to the credit card fraud detection problem, because they ignore the behavior patterns of the cardholder, and cannot deal with the nonlinear local boundary. In this paper, we propose a new method, MANE (Model-Agnostic Non-linear Explanations) for deep learning models, that can provide a set of interpretable features for each transaction. First, we use a cardholder's historical transactions to extract a set of behavior patterns for the cardholder via an aggregation strategy. Next, we use nonlinear model GBDT (Gradient Boosting Decision Tree) to mine cross features based on these behavior patterns, so that our explanation model can approximate the local boundary of a complex deep learning model. Finally, for every to-be-explained sample, we obtain its neighbors by using a random perturb method, then we train an interpretable linear regression model for the sample based on its neighbors, the behavior patterns of this cardholder and the related cross features. Our experiments on a real dataset demonstrate that our method achieves better effects.