Two‑Layer Dynamic Clustering and Trajectory Analysis of Gambling Behavior
Session Title
AI in Gambling: Mapping the Research Landscape & Advancing Behavioral Risk Detection
Presentation Type
Paper Presentation
Start Date
26-5-2026 12:00 AM
Abstract
Understanding how gambling behavior evolves over time is central to effective harm-prevention strategies, yet most existing approaches with unlabelled data rely on fixed clustering assumptions or cross-sectional snapshots. We introduce a two-layer dynamic clustering framework that integrates Gaussian Mixture Models with adaptive model selection and meta-clustering to track player behavior across consecutive time windows. Using six months of payment transaction data from a major gambling operator, we identify three key empirical patterns: (i) the number of behaviorally distinct player groups varies across periods, departing from the common fixed-cluster assumption; (ii) low-risk recreational types persist consistently, whereas medium- and high-risk profiles appear intermittently, producing a “long-tail plus extreme” structure; and (iii) risk escalation trajectories predominantly pass through medium-risk groups, underscoring their role as a gateway stage. By combining unsupervised learning with interpretable risk scoring and validating behavioral patterns against 12 months of wagering data with corresponding PGSI scores, our framework provides a reproducible basis for analyzing behavioral evolution and highlights medium-risk players as a targeted intervention window for regulators and operators.
Two‑Layer Dynamic Clustering and Trajectory Analysis of Gambling Behavior
Understanding how gambling behavior evolves over time is central to effective harm-prevention strategies, yet most existing approaches with unlabelled data rely on fixed clustering assumptions or cross-sectional snapshots. We introduce a two-layer dynamic clustering framework that integrates Gaussian Mixture Models with adaptive model selection and meta-clustering to track player behavior across consecutive time windows. Using six months of payment transaction data from a major gambling operator, we identify three key empirical patterns: (i) the number of behaviorally distinct player groups varies across periods, departing from the common fixed-cluster assumption; (ii) low-risk recreational types persist consistently, whereas medium- and high-risk profiles appear intermittently, producing a “long-tail plus extreme” structure; and (iii) risk escalation trajectories predominantly pass through medium-risk groups, underscoring their role as a gateway stage. By combining unsupervised learning with interpretable risk scoring and validating behavioral patterns against 12 months of wagering data with corresponding PGSI scores, our framework provides a reproducible basis for analyzing behavioral evolution and highlights medium-risk players as a targeted intervention window for regulators and operators.