Revisiting Alignment and Uniformity for Recommendation via Discrimination and Reliable Assessment

Xinzhe Jiang, Lei Sang, Shun Lian, Yi Zhang, Yiwen Zhang

Abstract: Utilizing alignment and uniformity for recommendation has shown success in considering similarities between users and items. Despite this effectiveness, we argue that they suffer from two limitations: (1) alignment loss as a measure of model quality fluctuates significantly during adjustment, leading to inaccurate assessments. (2) Current methods ignore potential connections for user-user and item-item, resulting in incomplete understanding of user preferences and item characteristics.To address these issues, we propose using the trace of user and item correlation matrices as a new assessment metric to replace traditional alignment for the first time. This design reduces the impact of hyperparameters on model assessment, ensuring that trace and model quality are optimized simultaneously, thereby improving recommendation accuracy. Based on this, we introduce a new model Alignment and Uniformity with Discrimination, which additionally considers the similarities for user-user and item-item. Specifically, DiscrimAU calculates the Euclidean distance between the user (item) relevance matrix and its fully aligned matrix, distinguishing the relevance levels among different users (items). This process ensures that highly relevant users and items are more closely aligned, capturing more information. Extensive experiments on three datasets show that the proposed model achieves a maximum improvement of 6.29%, clearly demonstrating its effectiveness.

External IDs:doi:10.1109/tce.2025.3527007