Linear Approximation to a Neural Network Model for E-Commerce Anomaly Detection

DOI: 10.5176/978-981-08-7466-7_kd-08

Authors: Reza Khosravani

Abstract:

Because of their superior performance over linear models, neural networks are widely used to detect irregular patterns in data. Unfortunately, non-linear models in general, and neural networks in particular, lack transparency and interpretability. In industries where a justification is needed for actions proposed by the model (e.g. customer credit management), a black-box approach for modeling is not desirable. In this paper, we propose a novel technique to convert a non-linear model to an interpretable linear model at the tail of the score range