"Personalized Graph-Based Retrieval for Large Language Models"

When AI lacks your history, it learns from your neighbors to personalize responses

PGraphRAG enhances LLM personalization by using graph-based retrieval instead of just user history, enabling better context even for new users with limited data .

-----

https://arxiv.org/abs/2501.02157

🤔 Original Problem:

Existing LLM personalization methods rely heavily on user history, making them ineffective for new users or those with limited interaction data. This creates a significant cold-start problem in real-world applications .

-----

🔧 Solution in this Paper:

→ PGraphRAG constructs bipartite graphs connecting users with items through interaction edges

→ The framework retrieves context using both direct user history and neighbor information from the graph

→ A query function transforms input into retrieval queries for finding relevant user profile entries

→ The system assembles personalized prompts by combining input with retrieved graph-based context

→ The framework optimizes retrieval by selecting the k-most relevant entries from user profiles

-----

💡 Key Insights:

→ Neighbor information provides significant value beyond just user history

→ Retrieval of 4 items generally yields optimal performance

→ The method works effectively even with limited user data

→ Graph-based approach enables richer context for cold-start scenarios

-----

📊 Results:

→ Outperformed baselines across 12 personalization tasks

→ Achieved +32.1% ROUGE-1 improvement in Hotel Experience Generation

→ Demonstrated consistent gains in both long and short text generation

→ Maintained performance with both BM25 and Contriever retrievers

------

Are you into AI and LLMs❓ Join my daily AI newsletter. I will send you 7 emails a week analyzing the highest signal AI developments. ↓↓

🎉 https://rohanpaul.substack.com/