Variational Autoencoders for Anomaly Detection in Choice Models

Abstract

Generative machine learning is increasingly gaining attention in the choice modelling field. By learning the underlying probability distribution of training data to synthesise new observations, these models offer a powerful framework for anomaly detection, i.e., identifying abnormal or atypical patterns in the input data. We introduce the use of variational autoencoders for anomaly detection. This is a generative model with two components, an encoder that compresses the input data into a low-dimensional latent representation, and a decoder that reconstructs the input from this latent representation to produce new data. By comparing the original data to reconstructions generated from samples of the latent space, reconstruction errors can be computed and used to flag atypical observations. We demonstrate that anomalies have a substantial impact on choice model outputs, such as values of time. This framework enables choice modellers to better detect outliers in high-dimensional datasets, thereby supporting data cleaning and model specification.