Mining Gambling Data for Modeling Gambling Behavior Patterns

Understanding player behavior for responsible gambling research is a difficult task due to the lack of data on players’ activities. Past studies in this area are largely limited to publicly available behavioral data or aggregated players data. Problem gambling in gamblers is typically identified only after they have already been addicted or have already been engaging in problematic gambling behavior. Furthermore, “risky” gambling behavior has historically been difficult to define due to the varying patterns of gambling activity that could potentially be attributed to it. In this dissertation we illustrate the methodology and algorithms used to engineer financial data for further analysis. We demonstrate the use of time-series analysis and anomaly detection using statistical and unsupervised machine learning methods to detect anomalous individual player behavior. This is accomplished using a custom metric (delta index) for statistical analysis and the spectral residual algorithm for machine learning based anomaly detection. We also demonstrate the use of longitudinal clustering using Gaussian Mixture Models to identify changes in player behavior at 30 day time intervals. This method groups players based on changes in behavior over time rather than their existing behavior in any given time slice. Additionally, we demonstrate the use of this model to track individual player behavior over time. The behavior analysis methods illustrated in this dissertation can be generalized for accepting additional behavioral features for further research in this area. Additionally, these methods can be further adapted for use in other behavioral studies.