基于TextRank的用户模板构建方法  

Method of Building User Profile Based on TextRank

作　　者：段准[1] 刘功申[1] 

机构地区：[1]上海交通大学信息内容分析技术国家工程实验室,上海200240

出　　处：《计算机技术与发展》2015年第10期1-6,共6页Computer Technology and Development

基　　金：国家自然科学基金资助项目(61472248;61171173)

摘　　要：在基于内容的推荐系统中,初始用户模板的准确性对后面的推荐精度有很大影响。因此,在系统初始时,必须从少量用户信息中准确地提取出用户兴趣模板,尽可能减少噪声的引入。否则会在后期更新模板时产生偏移性问题,造成推荐的不准确。针对此问题,文中提出了一种基于TextRank算法建立初始模板的方法。首先对所拥有的少量用户感兴趣文本进行预处理并确定词义项,然后进行聚类,接下来对聚类得到的每个类别分别以义项为单位构建TextRank模型,并引入相似度影响因子、共现度影响因子、类权重影响因子对TextRank模型中的概率转移矩阵进行改进。迭代之后选取每个类中最为关键的若干义项进行综合,得到最终的初始用户模板。实验结果表明,该算法得到的初始用户模板较为精确,可以达到较好的推荐效果。In content-based recommendation system, the accuracy of the initial user profile has a great influence on the accuracy of recom- mendation later. Therefore, profile must be built as precise as possible on condition of having little user information when the system is in initial state. Otherwise,it will bring offset when updating the user profile later, which will cause inaccuracy of recommendation. A method of building initial user profile based on TextRank is presented in this paper. At first, the texts user interested in are preprocessed and the meaning of each word is determined. Then, clustering operation is done and TextRank models are built by using meaning of word as unit. Various influence factors are also introduced to make the TextRank transition probability matrix better. At last, the most important mean- ings of word are chosen from each cluster to build the final initial user profile. Experimental results show that the accuracy of recommen- dation is high by using this method.

关 键 词：内容推荐算法 同义词词林 层次聚类 TextRank 图模型 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]