基于LDA主题模型和标签聚类的党建信息推送策略研究

【作者】 杨帆；

【导师】 施继红；

【作者基本信息】 云南大学 ， 控制工程， 2016， 硕士

【摘要】 互联网已成为人类生活不可缺少的工具,其双向互动功能强大的特点对传统的党建工作提出了新的挑战。政党必须建立起足够灵活的信息沟通渠道来宣传党的方针、政策,组织和动员广大党员群众。随着云南省党员队伍不断壮大,加快云南省网络党建的进程将是今后很长一段时期内基层党建工作的主题。党建信息推送是网络党建的重要应用需求。通过“网络党建”系统短／彩信发送平台完成一次全省200多万党员的党建信息全面推送需要大约12小时,无差别的党建信心推送耗时,浪费资源,且准确度低。因此,对党建领域内的信息推送策略进行研究,寻求合适的推送算法,是党建平台建设中的重要工作之一。云南省基层党建综合服务平台党建息采用短文本形式,缺乏足够的信息提供推送应用,同时针对平台现有推送服务扁平化、效率低等问题,本文借助隐含语义分析技术增强对文本内容语义层面的理解,采用LDA、Word2Vector以及聚类进行应用探索,提出了两种关注党员状态且符合实际要求的推送策略：1)利用LDA模型对用户历史数据作无监督聚类,根据党员反馈矩阵推测党员偏好,利用待推送信息与党员偏好的相关度实施过滤式党建信息推送；2)利用深度表示模型Word2Vec对党建手机报进行学习得到中文词向量,中文词向量归一化处理,之后运用／AP Cluster算法对词向量进行无监督聚类得到党建信息标签聚簇,考虑党员用户差异性,运用权值将党员与标签聚簇进行匹配,产生手机报定向推送结果。最后,本文以云岭先锋网手机报真实文本作为实验数据对本文两种推送算法进行了验证,采用召回率和准确度作为推送结果评价标准,结果表明相比传统协同过滤和无差别推送方法,两种推送策略跟踪党员偏好信息,考虑党员政治身份和社会关系其更符合实际需求,在取得了更稳定的推送效果的同时,实现了信息的组织和党员偏好的有效建模。更多还原

【Abstract】 Information pushing is an important application requirement of Network Party-construction. The traditional undifferentiated Information pushing of Yunnan Province grassroots party organizations integrated service platform was time consuming and resources wasting. In order to consider the individualized demand for commies, efficient and personalized Information pushing strategy is crucial to Network Party-construction in Yunnan Province.After analyzing the shortages of message sending platform of Yunnan Province grassroots party organizations integrated service platform, such as flattened serve, inefficiency, two personalized information pushing strategies for commies employing the LDA topic model, Word2Vector and Tag Clustering were presented in this paper.:1)By employing LDA topic model, unsupervised clustering was applied on the history data for Party members. Based on the feedback matrix, the Party members’ preference would be deduced, and then a message could be filtered and sent to the selected people considering the relevance between the message and the Party members’ preference.2) After splitting short messages into word segments, the Tags of short messages for Network party-construction could be transformed into vectors with Word2Vec toolset.; Then AP Cluster algorithm was applied on the vectors to cluster Tags by defining the Cosine distance between two vectors; By taking the Commies profile into account and measuring the importance of a tag Cluster to a Commies, an personalized information pushing strategy was presented. At last, as the test data, the real mobile news text on the website of Yunling Pioneer is input in the experimental tests and the results show that this two personalized strategies are more stable and can meet the needs of the real application comparing with the traditional collaborative filtering algorithm and undifferentiated pushing, meanwhile Information organizing and Modeling on Party members’ preference are also implemented.更多还原