Improving Pricing Recommendations Using Nearest Neighbors Retrieval via Contrastive Learning and Hard Negatives Mining

Abstract

Accurately determining selling prices for listings in online marketplaces poses a significant challenge due to the lack of universally recognized identifiers, such as Global Trade Item Numbers (GTIN) or Universal Product Codes (UPC). This lack of uniformity results in inconsistencies across product descriptions, titles, attributes, and features, complicating price prediction efforts. Traditional approaches for price prediction have predominantly relied on manually engineered features or direct price predictions from textual and image data, often failing to capture the nuanced differences between similar products. While transformer architectures have been widely used in e-commerce for item recommendation and retrieval tasks, these applications focus mainly on single-modal retrieval and do not address the complexities of pricing. In this paper, we introduce a novel approach to price recommendation by leveraging item retrieval methods enhanced with hard negatives during training. Incorporating hard negatives improves the quality of the generated embeddings, enabling more effective differentiation between similar listings with significantly different prices. This methodology focuses on understanding the contextual relationships and characteristics of listings relative to one another, rather than solely focusing on direct price prediction. By integrating contrastive learning with both price and aspects-based hard negatives, our approach better distinguishes between similar listings, significantly advancing price recommendation methods. Our research addresses this gap, aiming to significantly enhance the accuracy and effectiveness of pricing strategies. Extensive evaluations show that our method substantially improves pricing accuracy and enhances retrieval accuracy compared to existing approaches. We present extensive analysis and demonstrate successful deployment to production environment.