Fine-grained auxiliary learning for real-world product recommendation
This paper addresses the critical challenge of insufficient automated coverage in real-world recommender systems. We propose the Auxiliary Learning with Coverage (ALC) framework, which enhances model discrimination against hard negative samples via fine-grained auxiliary tasks, introduces a dual-objective optimization mechanism based on batch-wise hardest negatives, and incorporates a threshold-consistent margin loss to align similarity ranking with binary classification decisions. By integrating principles from extreme multi-label classification, ALC is evaluated on two large-scale benchmarks—LF-AmazonTitles-131K and Tech and Durables—demonstrating significant improvements in recommendation accuracy and stability under high-coverage scenarios. Experimental results show that ALC achieves state-of-the-art performance while maintaining high automated coverage, effectively meeting the practical deployment requirements of industrial-scale recommender systems.