Study of anomaly detection in diverse populations using probabilistic graphical models

Abstract

Most credit card transactions are characterised by randomly changing patterns of behavioural characteristics of the card users involved. The behavioural features tend to vary greatly due to the diversity inherent in most populations. In order to detect fraud and anomalies efficiently in these typical diverse populations using machine learning models, the models must be robust enough to capture the complex user behaviours in their variety. To address these challenges, we used base Hidden Markov Models (HMMs) and then a more richly expressive model called hierarchical Hidden Markov Model (HHMM) that can capture more dimensions and latent variables, which enables it to perform better. Furthermore, the model also reduces the learning times, which is very important when it comes to evaluating diverse population domain such as online credit card transactions as these impacts the response times. We evaluate the performance of HHMM in both individual and diverse population-based credit card transaction anomalies and labelled real-world timeseries data