Optimized Social Recommendations Using Graphs and Language Model Integration

The rapid growth of e-commerce and social-centric digital platforms has heightened the demand for personalized recommendations that account for both temporal user preferences and social relationships. Session-based Social Recommendation (SSR) models combine session-based dynamics with social interactions but face challenges in capturing personalized user information and maintaining computational efficiency. Existing SSR methods rely heavily on graph-based algorithms that, while effective, are computationally expensive during inference. To address these limitations, we propose LLM-BRec, a novel framework for efficient session-based social recommendations. Our approach employs a Social-aware Heterogeneous Graph (SHG) to integrate user-item interactions, item-side information, and social network data, enabling the learning of enhanced user and item representations. For session modeling, we leverage a BERT-based model, which provides faster inference by utilizing a self-attention mechanism to capture temporal dynamics within sessions. Additionally, Large Language Models (LLMs) are employed post-training to generate comprehensive user profiles, optimizing personalization while minimizing computational overhead. Experiments on both social and non-social datasets demonstrate that LLM-BRec significantly outperforms state-of-the-art SSR models in accuracy and efficiency, reducing inference time by $80 \%$ while maintaining robust recommendation performance. This work sets a new benchmark for session-based social recommendation by combining lightweight computation with advanced personalization.