Adaptive modeling combined with reinforcement learning to handle text multi-categorization tasks in news pushing
Abstract
With the swift advancement of information technology, news push—a pivotal means of information dissemination—has progressively evolved towards greater intelligence and personalization. User interest classification and precise push mechanisms, central to the development of news recommender systems, not only enhance the efficiency with which users access information but also effectively cater to their individualized needs. However, the performance of traditional methods in classification and recommendation remains constrained due to the high-dimensional nature of news data, limited annotations, and the dynamic fluctuations in user interests. In light of these challenges, this paper proposes BTQLM, a novel framework for news classification and recommendation that seamlessly integrates pre-trained language models with reinforcement learning. Initially, this study extracts contextual semantic features of the text through BERT, bolsters local feature extraction by incorporating a convolutional pooling layer, and captures global dependencies via the Transformer network. Building upon this, a reinforcement learning mechanism is introduced to dynamically refine the classification strategy by defining loss and reward functions, followed by joint optimization, thereby effectively enhancing classification performance. Furthermore, the paper combines the classification results with the NCF-based collaborative filtering method to enable personalized news recommendations through user interest modeling and similarity computation. Experimental results demonstrate that BTQLM significantly outperforms traditional methods such as TextCNN, ABLCNN-Word, Tree-LSTM, and AdaSent in classification accuracy, achieving over 0.85 accuracy on both the MIND and MIND-small datasets. In the recommendation task, the push strategy based on BTQLM’s classification results attains a recommendation accuracy exceeding 0.6 across diverse scenarios. The research presented herein offers an efficient technical solution for the design of intelligent news push systems, providing both theoretical underpinnings and practical guidance for advancing the field of news recommendation.