透過雙分群正則化改進單類別協同過濾模型

Abstract

雖然矩陣分解已經成為單類別協同過濾問題的主流方法，使用者之間的關係和項目之間的關係卻沒有被直接學習到。相似度計算在尋找使用者之間的關係和項目之間的關係上扮演著核心的角色。然而，由於回饋矩陣的稀疏度極高，在整條行為向量上做相似度計算會讓我們不容易為使用者和項目找到高品質的鄰居。為此，將雙分群的技術應用在回饋矩陣上來尋找使用者項目小群集是一個選項。然而，大部分雙分群的研究都在各個使用者項目小群集中執行局部並獨立的協同過濾模型，造成了排序導向的協同過濾模型無法學習到並未分類至同一個小群集的項目之間的喜好度差距。為了解決這個問題，我們提出了一個名為「雙分群正則化」的新架構，無縫地將著名的流形正則化和使用者項目分群結合在一起。相對於流形正則化，雙分群正則化同時降低了拉近帶雜訊鄰居的危險性以及計算開銷。實驗結果證明了雙分群正則化不但加強了矩陣分解中使用者之間的關係和項目之間的關係，也是一個能夠更佳地使用使用者項目分群來增進單類別協同過濾模型表現的方法。

Although Matrix Factorization (MF) has been the dominant approach in One-Class Collaborative Filtering (OCCF) problems, the user-user relationship and item-item relationship are not directly captured. The similarity computation plays the key role in discovering user-user relationship and item-item relationship. However, due to the high sparsity of feedback matrix, computing similarity regarding the entire behavior vector leads to the difficulty of finding high-quality neighbors of users and items. To this end, finding user-item subgroups by applying co-clustering techniques to the feedback matrix is an option. Nevertheless, most of the previous work applies a CF model locally and independently inside each discovered user-item subgroup, which makes ranking-oriented CF models fail to learn the preference differences between items which are not grouped into the same user-item subgroups. To deal with this problem, we propose a new framework Co-Cluster Regularization (CoReg), which seamlessly combines the well-known Manifold Regularization with user-item co-clusters. Compared to Manifold Regularization, CoReg simultaneously reduces the risk of drawing noisy neighbors and computation overhead. Experimental results show that CoReg not only reinforces the user-user relationship and item-item relationship of MF, but also serves as the better way to boost the performance of OCCF models by utilizing user-item co-clustering.