Enhancing Library Recommendation Systems with Integrated PSO for Parameter Tuning in BERT-Derived Multimodal Contexts

Abstract

This paper presents a novel multimodal recommendation system that integrates Particle Swarm Optimization (PSO) with Bidirectional Encoder Representations from Transformers (BERT) to address critical challenges in library reading promotion, including low recommendation accuracy and cold-start problems. The proposed system leverages multimodal data encompassing book text, cover images, and user behavior patterns. BERT facilitates deep semantic encoding of textual information, while PSO dynamically optimizes key hyperparameters including learning rate, batch size, and dropout rate, alongside multimodal fusion weights. Experimental validation was conducted using real-world data from a provincial public library. The PSO-BERT model demonstrated superior performance across all evaluated metrics: accuracy (0.831, +21.8% vs. collaborative filtering), recall (0.805), F1-score (0.818), and hit rate (0.852). User satisfaction surveys further confirmed significant improvements, with relevance and novelty scores reaching 8.6 and 7.9 points, representing increases of 21.1% and 33.9%, respectively, compared to traditional collaborative filtering. Ablation studies underscored the critical importance of multimodal feature integration, with the fused model maintaining F1-scores above 0.65, substantially outperforming unimodal configurations. The PSO-BERT integrated multimodal recommendation system exhibits substantial potential for enhancing recommendation accuracy, mitigating cold-start challenges, and improving user satisfaction in library reading promotion contexts.