A Machine Learning Approach to Social Work Intervention to Address High-Risk Problem Gambling

Background and Purpose: Since 2014, 38 states and the District of Columbia have opted to offer legal, online gambling or sports wagering opportunities. Compared to other states, those with more gambling offerings, particularly online, have up to three times the rate of gambling problems. Health, mental health, financial, and legal problems often co-occur with gambling problems, making this a pressing concern for social work. Individuals who bet at the highest intensity (e.g., allocation of time, money, and number of bets) represent a small population who gambling operators could target with harm-reduction interventions. Although most states require operators to offer voluntary limit-setting safeguards and other forms of self-help, there is, to date, no requirement that operators attempt to identify and intervene with players who are spending more than they can afford to lose. Protecting individuals and families from gambling-related harm will require a shift in policy from placing the responsibility for managing play solely on the individual to requiring that operators use information from their vast data sets to identify and address problem gambling behavior in their player population. The goal of this study is to use machine learning (ML) techniques with five years of online betting data from all operators in a Northeastern state to identify players gambling at increasingly higher risk levels; this will be used to guide voluntary or regulator-mandated early intervention and referral to services.

Methods: Unsupervised ML techniques were used to identify behavioral risk profiles from 8 million wager records from 128,547 individuals gambling online from 2016 through 2020. Feature engineering produced indicators of temporal dynamics, volatility, escalation, and anomalies. Highly correlated features (|r| > 0.9) were excluded, and dimensionality was reduced via Principal Component Analysis. K-means clustering was applied on the standardized feature set; the optimal number of clusters was selected using the elbow method and validated with silhouette analysis.

Results: Thirty-five variables, representing play patterns that could align with risk for problem gambling, were identified. The model achieved a strong silhouette score of 0.86, indicating high intra-cluster similarity and strong inter-cluster separation among three distinct groups. The high-risk cluster exhibited patterns such as sustained increases in betting, high volatility in daily betting behavior, and erratic bet size, which were further differentiated by behavioral markers.

Conclusions and Implications: These results demonstrate the utility of unsupervised learning to identify behavioral profiles and support the integration of clustering-based monitoring systems for early detection and intervention in online gambling platforms. We expect that our model, trained on data from a full range of operators in a jurisdiction, will be generalizable across sites in all states, allowing regulators to require increased accountability from operators. This, in turn, will contribute to broader policy and workforce development efforts to triage bettors at risk to a range of support and treatment services offered primarily by social workers. Implications on how social work organizational leaders can lead the development of a framework of services for at-risk individuals will be discussed.