How to "Improve" Prediction Using Behavior Modification

Galit Shmueli will explore how internet platforms leverage behavioral data to predict user behavior, both for internal purposes and for their business clients, such as advertisers, insurers, and governments.

Achieving high predictive accuracy is essential, and data scientists are continually developing advanced algorithms and utilizing more extensive and diverse datasets to improve it. However, these platforms can also boost prediction accuracy by subtly steering users' behavior to match predicted outcomes, using behavior modification techniques. This approach, often involving reinforcement learning algorithms, is not widely addressed in the machine learning and statistics literature. To thoroughly examine this strategy, the authors introduce causal reasoning, specifically Pearl's causal do(.) operator, into predictive modeling. They dissect the prediction error that arises when behavior modification is employed and discuss its implications for prediction accuracy, platform users, and business clients. While behavior modification can make predictions seem more accurate, it may not translate to effective real-world applications and could have unintended negative consequences for users.

Organized by:

Institute of AI in Management
LMU Munich