Excessive alignment hinders representation learning in cloud-scale collaborative filtering

Abstract Collaborative Filtering (CF) is a foundational paradigm in recommender systems, where recent work learns representations through two geometric principles: alignment, which pulls interacted user-item pairs closer, and uniformity, which spreads embeddings across the space. This problem is particularly relevant in modern cloud-scale settings, where models are trained on massive interaction data. While effective, existing approaches largely assume that stronger alignment is always beneficial. In this work, we provide a theoretical analysis showing that excessive alignment overly shrinks representation distances between users and their interacted items, leading to representation collapse. This collapse suppresses embedding diversity and expressiveness, ultimately limiting recommendation performance. To address this limitation, we propose a unified paradigm that balances alignment, uniformity, and separability, where separability encourages unrelated representations to remain distinguishable in the latent space.Experiments on four publicly available benchmark datasets demonstrate that our proposed method consistently outperforms state-of-the-art (SOTA) CF baselines. In particular, it achieves an average improvement of 3.52% in Recall@20 and 4.65% in NDCG@20 over the best performing baselines across all datasets. All theoretical proofs and additional experimental results are provided in the appendix, and the complete source code is included in the supplementary material.