Mr. Kailash Thiyagarajan

LLM-Based Aspect Augmentations for Recommendation Systems

Abstract:

Large Language Models (LLMs) have demonstrated remarkable effectiveness in various domains, including recommendation systems. This talk will explore how LLM-generated item aspects—which capture justifications for users’ purchase intentions—can enhance Learning-to-Rank (LtR) models. I will present a structured approach to aspect augmentation using feature concatenation, wide-and-deep models, and two-tower architectures.

The insights for this talk are drawn from recent research, particularly "LLM-Based Aspect Augmentations for Recommendation Systems" (Walmart Global Tech), which highlights the impact of LLM-generated aspects on ranking metrics such as MRR and NDCG. I will discuss how these techniques can be applied beyond eCommerce to search ranking, personalization, and other recommendation-based applications, offering practical takeaways for both academia and industry.