Game Intelligent Analyst - Anomaly Detection in Casino Games using Machine Learning Algorithms

Abstract

Anomaly detection is a very large and complex field. In recent years, several techniques based on data science were designed in order to improve the efficiency of methods developed for this purpose. In particular, density-based anomaly detection allows to estimate how much probability densities differ from each other by solving a supervised learning problem using data drawn from these densities as a dataset. Besides, the casino games company GAMING1 aims at automating the detection of behavior modifications in its games when an update is performed. This master thesis introduces a method based on density-based anomaly detection to reach GAMING1’s objective. We explore two main approaches to solve this problem. The first one relies on standard supervised learning algorithms. For this approach, two algorithms were considered: Extremely randomized trees and linear support vector machine. The second approach is based on deep learning and exploits recurrent neural networks architectures. Using experimental cases of anomalies in casino games, it’s been shown that extremely randomized trees outperforms both other methods as far as performances and visual interpretation are concerned, but also considering computational resources. In addition to accuracy, we rely on several other metrics related to the comparison of probability densities in order to assess formal performances of the presented algorithm. We show that the tree-based method allows us to distinguish irregular data with an accuracy of at least 0.7 for standard anomalies, while providing relevant visual support thanks to probability densities representation. Moreover, it’s also possible to identify more hardly detectable anomalies by using classifier calibration in order to enhance the visual support of the probability densities, despite an associated low accuracy.