A Head Start Matters: Dynamic-Calibrated Representation Alignment and Uniformity for Recommendations

The Bayesian personalized ranking (BPR) loss is a commonly used objective in training recommender systems, upon which various auxiliary graph-based self-supervised contrastive learning tasks are designed for improved model robustness. Previous research has also shown that the unsupervised contrastive loss shapes the learned representations from the perspectives of alignment and uniformity, and representations with lower supervised alignment and/or uniformity loss contribute to better model performance. Despite the progress, no one neither explores how the two representation qualities evolve along the learning trajectory, nor associates the behaviors with the combination of supervised and unsupervised representation alignment and uniformity (RAU). In this work, we first observe that different methods trades of alignment and uniformity to varying degrees, and hypothesize that optimizing over supervised RAU loss alone is not sufficient for an optimal trade-off. Then, by analyzing how BPR loss relates to the unsupervised contrastive loss where the supervised RAU loss stems from, we migrate the relation to propose our framework which aligns embeddings from both supervised and unsupervised perspectives while promoting user/item embedding uniformity on the hypersphere. Within the framework, we design a 0-layer embedding perturbation to the neural network on the user-item bipartite graph for minimal yet sufficient data augmentation, discarding the traditional ones such as edge drop. Extensive experiments on three datasets show that our framework improves model performance and quickly converges to user/item embeddings.