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/
