A deep reinforcement learning framework with hybrid actions and multi objective rewards optimizes E commerce personalized marketing

As the scale of e-commerce platforms continues to grow and user behavior becomes more dynamic, generating effective personalized marketing strategies has become essential for improving conversion efficiency and increasing long-term user value. Traditional approaches struggle with adapting to changing user states, modeling long-term returns, and supporting automated decision-making. To address these issues, this study develops a deep reinforcement learning framework for personalized marketing in e-commerce. The framework models user interactions as a Markov Decision Process and integrates user state representation, a hybrid action space, and a multi-objective reward design into a unified optimization process. The method adopts an improved Soft Actor-Critic algorithm to learn marketing strategies in an end-to-end manner. It introduces an adaptive temperature mechanism to balance exploration and exploitation, incorporates a constraint handling module to ensure business feasibility, and applies prioritized experience replay to improve sample efficiency and training stability. These modifications allow the model to better capture complex user behavior patterns and produce more reliable decisions. The experiments use an offline dataset that contains about 2.7 million real transaction and behavior records from an e-commerce platform. Compared with SAC-Base, SAC-IM improves CTR, CVR, and LTV by approximately 5.6%, 7.2%, and 8.3%, respectively. In addition, we further compare SAC-IM with advanced offline reinforcement learning baselines, including CQL, IQL, and TD3+BC, and report inference time, training time, and representative best/worst cases. These finding indicate that the method can balance short-term performance and long-term value, and demonstrate its effectiveness for large-scale e-commerce marketing tasks.