Summarizing ‘LLM-BRec: Personalizing Session-based Social Recommendation with LLM-BERT Fusion Framework’

BLOGS

Summarizing ‘LLM-BRec: Personalizing Session-based Social Recommendation with LLM-BERT Fusion Framework’

Tushar Prakash, Raksha Jalan, Onoe Naoyuki

30^th September 2024

Overview of proposed framework

Tushar Prakash summarises paper titled “LLM-BRec: Personalizing Session-based Social Recommendation with LLM-BERT Fusion Framework” co-authored by Raksha Jalan and Niranjan Pedanekar. Accepted at 2nd Workshop on Gen-IR at 47th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR)

Introduction

The rapid expansion of e-commerce and online entertainment has led to increased user expectations for personalized recommendations, driving the development of recommendation systems that effectively filter irrelevant information. Session-based Recommendation (SR) techniques focus on understanding user preferences within a temporal context to enhance prediction accuracy. Initial efforts in SR, such as Anonymous Session-based Recommendations (ASR), target scenarios where user IDs are not available, while Personalized Session-based Recommendations (PSR) utilize user IDs for improved cross-session information transfer. As social media gains traction, integrating social relationships into recommendation systems has become crucial, leading to the rise of Session-based Social Recommendations (SSR). Although SSR has shown promise through approaches like DGRec and SERec, they have limitations in fully leveraging personalized user information and tend to rely on computationally intensive algorithms.

To address these challenges, the proposed “LLM-BRec” framework introduces a Social-aware Heterogeneous Graph (SHG) for enhanced user and item representation and utilizes BERT for efficient session modelling. This framework builds a comprehensive knowledge graph from user interactions and social connections, allowing for better prediction of user interactions within sessions. By employing BERT’s self-attention mechanism, LLM-BRec significantly reduces training time by 50% and inference time by 80% compared to state-of-the-art methods. Furthermore, it emphasizes the importance of post-training user profiling with Language Models (LLMs) to enhance recommendation performance while maintaining computational efficiency. The effectiveness of LLM-BRec is validated through performance comparisons on both social and non-social recommendation datasets, consistently outperforming existing state-of-the-art models.

Key Results

Below are the key points derived from Tables 1 and 2 that highlight the superior performance of LLM-BRec:

Model Performance: LLM-BRec outperforms existing baselines (ASR, PSR, SSR) on both social and non-social datasets, showcasing its effectiveness in recommendation systems.
LLM-Based User Profiling: The significant improvement over traditional models highlight the importance of incorporating LLM-based user profiling to better understand user preferences.
Efficiency: LLM-BRec surpasses recent PSR and SSR models that utilize computationally heavy attention and graph-based algorithms for session modelling, indicating the effectiveness of bidirectional context in user-item interactions.
Social Network Utilization: While state-of-the-art models like DGRec and SERec leverage social networks for improved user preferences, they do not exceed LLM-BRec’s performance, demonstrating the power of LLM-based user profiling combined with SHG and BERT.
Comparison with RNN and LSTM: Experiments replacing BERT with RNN and LSTM show that BERT is more effective at creating rich representations of users’ session-level behaviour sequences due to its bidirectional context learning.
Non-Social SR Performance: LLM-BRec significantly outperforms other models (e.g., BERT4Rec, GRU4Rec, HRNN) in non-social SR scenarios, emphasizing the importance of efficient embeddings and capturing long-term user interests.
Overall Contribution: The results demonstrate that the combination of SHG for embeddings, efficient session modelling with BERT, and LLM-based user profiling enhances the overall performance of LLM-BRec.

Conclusion

This article presents LLM-BRec, a framework that enhances Session-based Social Recommendation (SSR) systems by utilizing LLMs for personalized user profiles, a Social-aware Heterogeneous Graph (SHG) for user and item representations, and BERT for session modelling. LLM-BRec improves recommendation accuracy while reducing computational costs, outperforming state-of-the-art methods across multiple datasets. Future research directions include integrating additional user context like temporal dynamics, exploring scalability in real-time systems, adapting the framework for various domains, incorporating multi-modal data sources, and enhancing user privacy in profiling.