"Unifying Generative and Dense Retrieval for Sequential Recommendation"

Smart fusion of retrieval methods makes recommendations work better with less memory

This paper introduces LIGER, a hybrid model that combines generative and dense retrieval methods for recommendation systems. It addresses performance gaps and cold-start item challenges while maintaining computational efficiency, showing significant improvements in recommendation quality on academic benchmarks.

-----

https://arxiv.org/abs/2411.18814

🔍 Original Problem:

Sequential dense retrieval models require storing unique representations for each item, leading to high memory costs. While generative retrieval offers a promising alternative, it struggles with cold-start items and shows performance gaps compared to dense retrieval methods.

-----

🛠️ Solution in this Paper:

→ LIGER integrates sequential dense retrieval into generative retrieval, using semantic IDs to represent items efficiently

→ The model combines text embeddings with semantic ID generation to improve cold-start item recommendations

→ During inference, LIGER uses beam search to retrieve candidates, then ranks them using dense retrieval techniques

→ The hybrid approach maintains the storage efficiency of generative retrieval while leveraging the ranking capabilities of dense retrieval

-----

💡 Key Insights:

→ Generative retrieval methods tend to overfit to items seen during training

→ Dense retrieval excels at cold-start items but requires significant storage

→ Semantic IDs effectively capture item relationships while reducing storage needs

→ Hybrid approaches can balance performance and computational efficiency

-----

📊 Results:

→ LIGER achieves 13.06% Recall@10 on cold-start items vs 0% for baseline TIGER

→ Storage complexity reduced from O(N) to O(t), where t << N

→ Performance matches or exceeds state-of-the-art on both in-set and cold-start items