協同過濾的中心點正規化

Abstract

協同過濾（CF）在推薦系統的進步中扮演著關鍵角色，通常涉及三個主要組成部分：編碼器、損失函數和負採樣。現有的研究往往強調設計複雜的編碼器來捕捉更高階的相似性，而忽略了損失函數的影響。在本文中，我們介紹了一種名為「中心點正則化」（CentReg）的新方法，旨在根據用戶和項目的潛在社群結構來調整嵌入向量的長度，進一步增強推薦效果。我們的研究首先發現了DirectAU在嵌入向量長度的潛在問題，然後設計出 CentReg，以正則化的方式輔助 DirectAU，且不與原先的最佳化目標互相衝突。我們對多個數據集進行的全面評估，突出了CentReg在推薦表現和計算效率方面，為對比性損失函數如 DirectAU 在 CF 任務上帶來的提升。

Collaborative filtering (CF) is pivotal to the advancement of recommender systems, typically involving three key components: the encoder, loss function, and negative sampling. Existing research often emphasizes designing complex encoders to capture higher orders of proximity while overlooking the impact of the loss function. In this paper, we introduce Centroid Regularization (CentReg), a novel approach aimed at enhancing recommendations by adjusting embedding magnitudes in CF through the underlying community structures of users and items. Our study first identifies a potential issue with DirectAU related to embedding magnitudes and then designs a regularizer called CentReg to address this issue, without conflicting with the original optimization. Our comprehensive evaluation across multiple datasets highlights the recommendation performance and the training efficiency of CentReg in enhancing contrastive loss methods, such as DirectAU, for collaborative filtering tasks.