基于改进张量分解模型的个性化推荐算法研究  

Personalized Recommendation Algorithm Based on Modified Tensor Decomposition Model

作　　者：陈梅梅[1] 薛康杰 Chen Meimei Xue Kangjie(Glorious Sun School of Business ＆ Management, Donghua University, Shanghai 200051, China)

机构地区：[1]东华大学旭日工商管理学院,上海200051

出　　处：《数据分析与知识发现》2017年第3期38-45,共8页Data Analysis and Knowledge Discovery

基　　金：国家社会科学基金项目"中国特色的网络消费调查研究"(项目编号:10BGL027)的研究成果之一

摘　　要：【目的】在基于张量分解的个性化推荐中,解决因UGC标签冗余、热门标签和资源影响用户个性化兴趣所导致的推荐准确性降低问题。【方法】提出一种改进的基于张量分解模型的个性化推荐算法,引入标签综合共现结合谱聚类的方法,借鉴TF-IDF中IDF的思想提出一种基于共现标签和资源的热门惩罚机制,对基于<用户,标签簇,资源>三元关系的初始张量进行重新定义。【结果】基于Last.fm数据集的仿真实验结果表明,从准确率、召回率和F1值各项指标上,本文提出的算法均有良好表现,综合共现谱聚类的引入使得推荐算法在F1值上平均提升5.91%,基于IDF改进初始张量后的推荐算法在F1值上平均提升1.29%。【局限】未针对其他领域的数据集进行验证,如微博、Delicious等。【结论】基于改进的张量分解模型的个性化推荐算法能够显著提高准确性,有利于社交网络环境下提供更令用户满意的资源。[Objective] This paper tries to improve the prediction accuracy of personalized recommendation algorithm based on the tensor decomposition model. [Methods] First, we proposed a new tensor model using spectral clustering technique based on combined tag co-occurrence. Second, we established a penalty scheme on popular tag and resource co-occurrence with the help of IDF in TF-IDF. Finally,we re-defined the initial tensor on the triplets of user, tag cluster, and resource. [Results] We examined the proposed model with dataset from Last.fm and found its precision, recall and F1 measure outperformed other algorithms. The F1 measures were increased by 5.91% and 1.29% thanks to the two proposed modifictions based on clustering and IDF. [Limitations] The proposed algorithm should be further evaluated with datasets from Weibo, Delicious, and other resources. [Conclusions] The new algorithm based on advanced tensor decomposition model could significantly improve the accuracy of resources recommendation to satisfy social network system users＇ information needs.

关 键 词：个性化推荐 UGC 标签 标签共现 谱聚类 张量分解 

分 类 号：TP391.3[自动化与计算机技术—计算机应用技术]