Leveraging LLMs and Heterogeneous Knowledge Graphs for Persona-Driven Session-Based Recommendation

A team of researchers has introduced a new AI framework designed to solve a core problem in e-commerce and streaming: making personalized recommendations for users who are anonymous or have sparse interaction history. Traditional session-based recommendation systems (SBRS) struggle with personalization because they only see short-term click sequences without user identity. This new method, detailed in the paper "Leveraging LLMs and Heterogeneous Knowledge Graphs for Persona-Driven Session-Based Recommendation," tackles this by constructing a rich, structured knowledge graph. This graph integrates user-item interactions, item features, and metadata from sources like DBpedia, which is then initialized with rich item embeddings generated by a Large Language Model (LLM).

In the first stage, the system uses an unsupervised learning technique called Heterogeneous Deep Graph Infomax (HDGI) to infer latent user personas directly from this knowledge graph, bypassing the need for personal data. In the second stage, these learned persona representations are combined with the LLM's understanding of items and fed into a modified sequential recommendation model. This creates a candidate set of items, which is then re-ranked to emphasize the user's immediate session intent. The result is a system that grounds user modeling in structured relational signals rather than just anonymous click sequences.

Experiments on major Amazon datasets for books and movies showed that this LLM-KG hybrid approach consistently outperforms standard sequential models that rely solely on session history. By effectively creating a "persona" for an anonymous user based on the interconnected web of items and their attributes, the framework offers a path toward more intelligent, context-aware, and personalized recommendations in privacy-sensitive or cold-start scenarios where little user data is available.