Contrastive Learning for Diversity-Aware Product Recommendations in Retail

TL;DR

The paper tackles the long-tail challenge in large-scale retail recommender systems by enhancing catalog exposure through contrastive learning integrated into IKEA's embedding-based pipeline. It reframes training to predict product embeddings and evaluates two negative-sampling strategies (in-batch and adaptive top-$k$) on top of a cosine-similarity objective, showing improved catalog coverage and diversity with minimal or no loss in ranking quality. Across IKEA Netherlands and RetailRocket datasets, offline results demonstrate substantial gains in coverage and Gini, with competitive NDCG@10, and online experiments confirm gains in exposure and diversity, albeit with dataset-dependent best strategies. This approach offers a practical pathway to richer catalog discovery in production pipelines without compromising recommendation accuracy.

Abstract

Recommender systems often struggle with long-tail distributions and limited item catalog exposure, where a small subset of popular items dominates recommendations. This challenge is especially critical in large-scale online retail settings with extensive and diverse product assortments. This paper introduces an approach to enhance catalog coverage without compromising recommendation quality in the existing digital recommendation pipeline at IKEA Retail. Drawing inspiration from recent advances in negative sampling to address popularity bias, we integrate contrastive learning with carefully selected negative samples. Through offline and online evaluations, we demonstrate that our method improves catalog coverage, ensuring a more diverse set of recommendations yet preserving strong recommendation performance.

Figures (2)

• Figure 1: The long-tail of item popularity as captured in our recommendations for the Netherlands, in time period $t$.
• Figure 2: The main algorithmic components of the existing recommender system at IKEA Retail.