"LLM is Knowledge Graph Reasoner: LLM's Intuition-aware Knowledge Graph Reasoning for Cold-start Sequential Recommendation"

LLMs guide Knowledge Graphs to make smarter recommendations with limited user data.

LIKR combines LLMs with Knowledge Graphs through reinforcement learning to solve cold-start recommendation challenges by treating LLMs as intuitive path reasoners.

-----

https://arxiv.org/abs/2412.12464

🤔 Original Problem:

→ Traditional recommendation systems struggle with cold-start scenarios where user interaction data is limited

→ LLM-based recommendations face scalability issues due to token limits

→ Knowledge Graph methods lack temporal awareness and perform poorly with sparse data

-----

🔧 Solution in this Paper:

→ LIKR treats LLMs as KG path reasoners that output intuitive exploration strategies.

→ The system feeds temporally-aware prompts to LLMs to predict user preferences.

→ A reinforcement learning agent explores the KG using rewards from both LLM intuition and KG embeddings.

→ The model combines LLM's general knowledge with KG's domain-specific insights through carefully balanced reward functions.

-----

💡 Key Insights:

→ LLMs can effectively guide KG exploration without needing the entire dataset

→ Temporal awareness in prompts significantly improves recommendation quality

→ Optimal balance between LLM intuition and KG embedding rewards varies by domain

-----

📊 Results:

→ Outperforms state-of-the-art methods on MovieLens-1M with higher recall@20 and nDCG@20

→ GPT-4-preview shows best performance among tested LLMs

→ Achieves 4.83% recall and 19.14% nDCG on MovieLens-1M dataset