Lessons Learned from Problem Gambling Classification: Indirect Discrimination and Algorithmic Fairness

Christian Percy, Artur S. d'Avila Garcez, Simo Dragicevic, Sanjoy SarkarAAAI 2020

Abstract

This paper examines the tension between data privacy and algorithmic fairness in the identification of at-risk gamblers. Using real-world data from two gambling operators, the authors investigate the risks of indirect discrimination in "gender-blind" models. They propose a "model-matched" technique to detect proxy-based bias and a "blind-separate" ensemble approach that improves fairness across genders without requiring sensitive data collection during live deployment.

Introduction

Machine Learning (ML) is increasingly utilized by gambling operators to identify players at risk of harm. To comply with privacy regulations like GDPR and avoid direct discrimination, these models are often designed to be "gender-blind," excluding gender as an input variable. However, this approach carries two significant risks: indirect discrimination, where the model learns gender through proxy variables (e.g., betting patterns), and algorithmic unfairness, where a single model performs with varying accuracy across different demographic groups. This study seeks to improve fairness—specifically parity in the True Positive Rate (TPR)—without forcing users to provide sensitive demographic data at the point of intervention.

Methodology

The authors introduce two novel technical contributions to address these challenges:

  • Model-Matched Indirect Identification: To determine if a model is "secretly" identifying gender through complex, non-linear relationships, the authors trained a secondary model using the same architecture and features specifically to predict gender. If this model's performance remains near random chance, it suggests the original model is truly gender-blind.
  • Blind-Separate Models (Ensemble Approach): The authors developed an ensemble system. During training, gender data is used to build three specialized models: one for confirmed male players, one for confirmed female players, and one for unspecified users. During deployment, the system remains "blind" by running all three models on every player. Following a "precautionary principle," if any of the three models flags a player as at-risk, an intervention is triggered. This allows the system to leverage gender-specific behavioral insights without requiring the user's gender at runtime.

Data and Baseline Performance

The study utilized real-world data from 22,500 players across two distinct gambling operators: a slots-focused brand (Operator 1, n=4,340) and a bingo-focused brand (Operator 2, n=18,275). The models utilized approximately 40 behavioral and transactional variables to predict voluntary self-exclusion, which served as a proxy for gambling harm.

Table 1: Baseline metrics by gender for Operator 1 (slots-focused) and Operator 2 (bingo-focused), showing disparities in True Positive Rates (TPR) and training data balance.

As shown in Table 1, the baseline performance varied significantly between operators. Operator 2 exhibited a notable "pro-female" bias, with a TPR for women (53.7%) that was 7.2 percentage points higher than for men (46.5%). In contrast, Operator 1 showed much smaller disparities, suggesting that algorithmic fairness is highly dependent on the specific platform and user base.

Key Results and Insights

  • Detection of Proxies: The "model-matched" test indicated low indirect discrimination; gender could not be reliably predicted from behavioral data, suggesting the baseline models were not heavily influenced by gender proxies.
  • Fairness Improvement: The "blind-separate" ensemble successfully reduced the gender gap in TPR for Operator 2 from 7.2% to 4.0%.
  • The Precautionary Trade-off: Improving fairness through the ensemble approach came at a cost to specificity. The True Negative Rate (TNR) dropped (e.g., from 98.1% to 91.9% for men in Operator 2), meaning more players were flagged who might not have been at risk. The authors argue this is justified in a public health context where the cost of a "False Negative" (missing a person in harm) is higher than a "False Positive."

Conclusion and Future Work

Simply removing a protected variable does not guarantee fairness and can often hide disparities in how well a model protects different groups. While the "blind-separate" approach offers a path toward more equitable harm detection, it requires balancing sensitivity against the risk of over-intervention. Future work will focus on establishing sector-wide standards for "acceptable" levels of bias and exploring how these models perform for casual or early-stage players compared to the regular players studied here.

Summary updated

BibTeX
@inproceedings{percy2020lessons,
  author = {Christian Percy and Artur S. d'Avila Garcez and Simo Dragicevic and Sanjoy Sarkar},
  title = {{Lessons Learned from Problem Gambling Classification: Indirect Discrimination and Algorithmic Fairness}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
}
Figure