融合主题模型及多时间节点函数的用户兴趣预测研究  

User Interest Prediction Combing Topic Model and Multi-time Function

作　　者：桂思思[1] 陆伟[1,2] 黄诗豪 周鹏程[1] 

机构地区：[1]武汉大学信息管理学院,武汉430072 [2]武汉大学信息资源研究中心,武汉430072

出　　处：《现代图书情报技术》2015年第9期9-16,共8页New Technology of Library and Information Service

基　　金：教育部人文社会科学基地重大项目"面向细粒度的网络信息检索模型及框架构建研究"(项目编号:10JJD630014);国家自然科学基金面上项目"面向词汇功能的学术文本语义识别与知识图谱构建"(项目编号:71473183)的研究成果之一

摘　　要：【目的】针对用户兴趣随时间推移不断变化的问题,利用主题模型及时间节点函数预测用户兴趣。【方法】使用主题模型生成用户兴趣,针对用户的所有兴趣,分别利用多时间节点函数对每个兴趣的每次出现进行加权,用以预测用户兴趣在下一个时间节点的分布情况。【结果】在Sogou搜索日志上,与基于记忆的用户兴趣模型、基于遗忘曲线的用户兴趣度多阶段量化模型进行对比实验,余弦相似度及KL(Kullback-Leibler)距离均表明本文方法能较准确地预测用户兴趣。【局限】仅在Sogou搜索日志上进行实验测试,还需在其他数据集上进一步检验。【结论】充分考虑用户历史数据中每一个时间点可更准确地对用户兴趣进行预测。[Objective] User interest is not static and it changes dynamically as time goes by, this paper proposes a user interest prediction model based on topic model and multi-time function. [Methods] Generate user interests by topic model, and calculate the weights of each user interest at every time point by applying multi-time function in order to predict user interest at next time point. [Results] Compared with memory-based user profile model and multi-step user profile model, cosine similarity and Kullback-Leibler divergence of the experimental results on search engine log data provided by Sogou Lab show that this model can predict user interests more effectively. [Limitations] The proposed method is only tested on search engine log data provided by Sogou Lab, and it need further examination on other data sets. [Conclusions] It is more effective to take every time point of user history data into consideration for user interest prediction.

关 键 词：主题模型 时间函数 用户兴趣预测 

分 类 号：G252[文化科学—图书馆学] TP391.1[自动化与计算机技术—计算机应用技术]