带有隐式反馈的SVD推荐模型高效求解算法

Efficient solution of the SVD recommendation model with implicit feedback

作为推荐系统的重要组成部分,协同过滤已成为了当今主流的推荐方法之一.其中基于潜在因子的协同过滤常采用SVD推荐模型分析用户喜好.近年来,随着SVD推荐模型研究的深入,SVD++,TrustSVD等一类带有隐式反馈的SVD推荐模型被相继提出.此类模型能更有效地从有限的数据源中挖掘有用信息并取得了较好的效果,因此受到了人们广泛关注.然而,现有文献大多关注于模型设计,缺乏专门针对带有隐式反馈的SVD推荐模型的高效求解算法.为此,本文首先研究了一般性的SVD推荐模型梯度求解框架,然后以SVD++推荐模型为突破口,基于块梯度下降方法设计了高效求解算法BCDSVD++并解决了容量矩阵求逆、稀疏数据优化处理等两个关键问题.实验表明,本文所设计的BCDSVD++算法具有比传统的并行梯度下降法更高效的求解效率及收敛能力.

Collaborative filtering,an important part of the recommendation system,has become a mainstream recommendation method.In collaborative filtering methods based on potential factors,SVD recommendation models are often used to analyze user preferences.With the recent research of SVD recommendation models,some SVD recommendation models with implicit feedback,such as SVD++and Trust SVD,have been successively proposed.These types of models can more effectively mine useful information from limited data sources and achieve better results than traditional SVD recommendation model,thereby garnering widespread attention.However,most existing papers focus on model design and the lack of efficient algorithms for SVD recommendation models with implicit feedback.Therefore,this paper first studies the general gradient solution framework of the SVD recommendation model.Then,it considers the SVD++recommendation model as a breakthrough and designs an efficient solution algorithm,namely,BCDSVD++,based on the block coordinate descent method.Furthermore,we solve the two key problems of capacity matrix inversion and sparse data optimization processing.Experiments show that the proposed BCDSVD++algorithm yields better solution efficiency and convergence ability than the traditional parallel gradient descent method.