Reinforcement Learning in Recommender Systems for Personalized Financial Decision Support: A Systematic Mapping Study

Reinforcement learning (RL) and bandit-based methods enable recommender systems to adapt to user and environmental feedback, offering opportunities for personalized financial decision support. However, evidence concerning their methodological integration, decision formulation, and evaluation remains fragmented across application domains. This systematic mapping study examines empirical RL-enabled recommender systems across four dimensions: RL approach, integration role, decision formulation, and evaluation practices. A systematic search of Scopus, ACM Digital Library, ScienceDirect, SpringerLink, and IEEE Xplore identified 7531 records covering peer-reviewed publications from 2015 to 2025. Following deduplication, screening, eligibility assessment, and quality appraisal, 50 studies were retained, of which 28 underwent focused empirical mapping. Bandit-based methods were the most prevalent approaches (39.3%), followed by value-based RL (32.1%), policy-gradient and policy-optimization methods (14.3%), model-based planning and tree search (7.1%), actor–critic methods (3.6%), and mixed value-based and policy-optimization RL (3.6%). RL-related mechanisms supported direct recommendation and policy optimization, recommender-strategy selection, representation or model adaptation, and preference elicitation or planning. Reward formulations encompassed behavioral responses, recommendation utility, economic objectives, operational costs, and feedback-derived preferences. However, direct financial-product recommendation applications were limited, and evidence linking adaptive-learning performance to sustained user-level financial outcomes remained insufficient. Although some studies incorporated online or human-participant evaluation, longitudinal validation of personalized savings and investment outcomes was limited. The findings emphasize the need to align RL mechanisms with financial decision contexts, user representations, reward functions, feedback structures, and suitability constraints. This study provides a structured methodological mapping and identifies financially meaningful reward design, constraint-aware adaptation, and longitudinal evaluation as priorities for advancing personalized savings and investment decision support.