Fine-grained auxiliary learning for real-world product recommendation

Date

October 2025

Source

SEPLN’25. XLI International Congress of the Spanish Society for Natural Language Processing (Journal paper)

Authors

Mario Almagro
Diego Ortego
David Jiménez

Summary of Our Research

Boosting Coverage in Product Recommendation with Auxiliary Learning
Product recommendation involves retrieving the closest items to a given query from a large product catalog, where queries must be matched with multiple relevant items. While widely studied, real-world deployment faces a critical challenge: ensuring high coverage, meaning most recommendations are automated without manual review.

Our research introduces ALC (Auxiliary Learning for Coverage), a strategy that learns fine-grained embeddings to improve coverage. ALC uses two novel training objectives that exploit the hardest negatives in each batch, creating stronger signals to distinguish relevant from irrelevant products. Tested on two large-scale datasets, LF-AmazonTitles-131K and Tech & Durables (proprietary), ALC achieves state-of-the-art coverage when combined with a threshold-consistent margin loss, paving the way for more scalable and accurate recommendation systems.