Reinforcement Learning-Based Adaptive Recommendation Systems for Real-Time Personalization in E-Commerce Platforms

Abstract

The objective of this study was to develop and empirically evaluate a reinforcement learning–based adaptive recommendation system capable of optimizing real-time personalization and improving recommendation effectiveness, user engagement, and conversion performance in a large-scale e-commerce platform environment. This study employed a quantitative, applied, and computational research design using real-world behavioral and transactional data collected from 12,480 active users of a major e-commerce platform in Tehran over a six-month period in 2025. User interaction data included clicks, views, add-to-cart events, purchases, and contextual browsing information. The recommendation problem was modeled as a Markov Decision Process, where the recommendation system functioned as an intelligent agent interacting dynamically with user environments. A Deep Q-Network reinforcement learning model was developed and trained using historical interaction sequences, experience replay, and epsilon-greedy exploration strategies. Model performance was evaluated using multiple recommendation effectiveness metrics, including click-through rate, conversion rate, precision@10, recall@10, normalized discounted cumulative gain, cumulative reward, and user engagement indicators. Comparative analysis was conducted against baseline models, including collaborative filtering, matrix factorization, and static ranking approaches. The reinforcement learning–based recommendation system demonstrated statistically and practically significant improvements across all performance metrics compared to conventional recommendation models. The proposed model achieved a click-through rate of 10.96% and a conversion rate of 4.87%, representing substantial improvements over baseline approaches. Precision@10 and recall@10 increased to 0.312 and 0.284, respectively, while normalized discounted cumulative gain reached 0.347, indicating superior recommendation ranking quality and relevance. The model exhibited stable convergence behavior, with cumulative reward increasing progressively and loss values decreasing significantly across training iterations. Additionally, user engagement indicators improved substantially, including increases in session duration, click frequency, add-to-cart rate, and purchase completion rate, confirming enhanced recommendation effectiveness and behavioral impact. The findings confirm that reinforcement learning provides a highly effective framework for real-time adaptive personalization in e-commerce recommendation systems by enabling continuous learning, dynamic optimization, and context-aware recommendation strategies. The reinforcement learning–based system significantly outperformed traditional recommendation methods in terms of recommendation accuracy, engagement, and conversion performance. These results highlight the potential of reinforcement learning to transform digital commerce personalization by improving user experience, optimizing recommendation effectiveness, and enhancing platform performance in dynamic and data-intensive environments.