ReAU: A Global-to-Local Perspective for Refining Alignment and Uniformity in Collaborative Filtering

Collaborative filtering (CF) plays a central role in recommender systems. Recent advances have inspired a series of paradigms, as alignment and uniformity (AU), which explicitly characterize the distribution of user/item representations over the unit hypersphere. By shaping representations to be both discriminative and well-spread, AU helps improve space distribution under sparse user–item interactions. Despite their effectiveness, this paradigm still has limitations: (i) over-emphasizing representation properties while overlooking the optimization dynamics leads to distorted representation distributions; and (ii) directly optimizing these objectives can be overly simplistic, failing to capture complex collaborative relations and inducing the optimization bias. In this paper, we revisit alignment and uniformity from global to local perspective in CF. The notion of local space is then introduced, followed by formal definitions of local alignment and local uniformity. Additionally, we theoretically reveal the optimization mismatch between global and local AU objectives, thereby biasing optimization toward dense space. Motivated by these insights, we propose a novel representation learning framework, named ReAU. It adopts multi-term local alignment to capture richer collaborative signals, and leverages rebalanced local uniformity to encourage more dispersed distributions. Extensive experiments on four public datasets show the effectiveness of ReAU.