Adaptive Attention-based State Representation in reinforcement learning based recommendation systems

Traditional recommendation systems often grapple with data sparsity, high dimensionality, and dynamic user preferences. While deep learning(DL)-based recommendation models excel at learning static user preferences through powerful representation learning, they struggle to capture the diversity and temporal dynamics of user interests. reinforcement learning(RL)-based on Markov decision processes(MDPs) offers distinct advantages by modeling evolving user–environment interactions, enabling dynamic decision-making for both short-term and long-term recommendation objectives. In deep reinforcement learning(DRL)-based recommendation systems, the agent must perceive current user states, process historical interactions, and make future-oriented decisions. A critical challenge lies in deriving effective state representations from user features and historical interactions. Existing approaches employ CNN or RNN models for RL state representation but often fail to capture fine-grained contextual signals. Although attention mechanisms enhance the model’s ability to track dynamic user interests, they suffer from instability and local optima. Toaddress these challenges, we propose an Adaptive Attention-based State Representation framework, the first to integrate adaptive attention with reversible instance normalization for RL state representation. It generates robust and precise state embeddings, significantly improving policy learning in DRL and demonstrating superior performance in cumulative rewards and user engagement metrics.