Interpretable behavioral clusters of gamblers through unsupervised learning.

Understanding the heterogeneity among highly involved gamblers is critical for the development of effective harm reduction strategies. This study employs unsupervised machine learning to segment a population of high-intensity Electronic Gambling Machine (EGM) users based on behavioral indicators derived from transactional data. Using a combination of Uniform Manifold Approximation and Projection (UMAP) for nonlinear dimensionality reduction and Density-Based Spatial Clustering of Applications with Noise (DBSCAN) for cluster identification, we performed a systematic grid search to optimize internal validity metrics (Silhouette = 0.5827, Davies-Bouldin Index = 0.4442, Calinski-Harabasz Index = 8907.46). The analysis yielded four well-separated behavioral clusters: one marked by impulsive withdrawals and night-time play; another showing consistent and high-frequency gambling; a third characterized by structured, high-stakes sessions; and a fourth exhibiting rapid, binge-like activity within short time windows. To facilitate interpretation, we trained cluster-wise random forest classifiers, identifying key discriminative features such as balance trajectory, inter-session timing, and variability in transaction intervals. Our findings demonstrate that high involvement is not a uniform construct, but rather encompasses diverse behavioral subtypes, each potentially associated with different levels of gambling-related harm. This segmentation framework offers practical implications for personalized responsible gambling initiatives and contributes to ongoing research advocating for data-driven player protection strategies.