Online Platform user Behavior Prediction and Decision Optimization based on Deep Reinforcement Learning

With the rapid development of Internet technology, online platforms have become an indispensable part of daily life, generating a large amount of user data in various fields such as shopping, social networking and entertainment. Effectively mining this data to provide personalized services has become an important focus in both academia and industry. The latest advances in deep learning and reinforcement learning have driven innovation in recommendation systems, particularly in addressing data sparsity and cold start issues. However, existing methods still struggle to capture the dynamic changes in user interests and fully consider the long-term impact of historical and future data. To address these challenges, this paper proposes two recommendation models based on deep reinforcement learning, word embedding techniques, and bidirectional LSTM. The first model combines GloVe word embedding with bidirectional LSTM to integrate historical and future data, improving decision accuracy. The second model introduces BERT for deeper user behavior modeling and uses principal component analysis to optimize the candidate set of items, thereby enhancing dynamic and comprehensive recommendations. The experimental results of the Amazon Music and Beauty datasets show that both models outperform traditional baseline models, with models based on BERT and bidirectional LSTM exhibiting excellent performance.