基于社交网络的个性化微博关注推荐系统的研究与实现

发布时间：2018-04-15 17:20

本文选题：微博关注推荐 + 社交相似度　；参考：《山东大学》2017年硕士论文

【摘要】：随着大数据时代的到来,新技术层出不穷,社交网络的发展如火如荼。微博是最热门的社交平台之一,拥有着庞大的用户群体,每天产生无数热点信息。在微博中,人们可以发布原创消息;用户可以在系统中找出自己感兴趣的对象,成为其粉丝;转发、评论、@等行为极大地丰富了用户之间的互动体验,也使得微博用户之间的交互更加多元化。然而,信息的泛滥也让用户难以选择,出现了信息过载的现象。推荐系统是用户和项目之间的桥梁,能够挖掘和捕捉用户的偏好,主动给用户推荐相关内容,目前已经被应用在很多场景下。协同过滤算法是其中最为经典的算法之一,然而该算法非常依赖用户-项目之间的评分数据,并且面临着严峻的数据稀疏性问题。在微博中,不存在用户对于项目的评分数据,因此不能简单地将协同过滤算法应用在微博关注推荐中。微博的社交网络特征给推荐问题提供了更多解决方案,融入社交行为、社交信任、邻居意见、隐语义模型等都会大大改善推荐的性能。本文首先对推荐系统的发展以及微博关注个性化推荐进行了研究,介绍了协同过滤算法的相关技术和原理,阐述了当前算法面临的困难与挑战。通过腾讯微博数据集分析了微博社交网络的相关特征、社交图谱、用户关系等,重新定义了微博关注推荐的相关术语,对微博中的不同社交行为进行建模,并介绍了系统的整体流程、技术平台、系统环境等。针对Top-N推荐问题,提出了基于社交相似度的微博关注Top-N推荐算法。根据微博关注行为、互动行为以及历史推荐记录分别计算相似度,通过计算出来的相似度找出最近邻集合,在此基础上给用户进行推荐。在微博数据集上对比了不同相似度计算方法的准确率、召回率和Fl-measure,并在Hadoop平台上利用MapReduce对算法进行了并行化设计,提高了算法的执行效率。针对评分预测问题,提出了融合社交信任和隐语义模型的微博关注推荐算法。将用户的历史推荐记录建模为评分矩阵,引入社会化推荐,通过用户之间的互动行为数据(包括@、评论和转发)计算用户之间的隐式信任,从用户的直接社交关系中得到用户之间的显式信任,将显式信任和隐式信任结合来构建扩展信任矩阵并融入SVD++模型。最终在KDD Cup 2012数据集上的实验表明算法在RMSE和MSE上得到了更好的结果。
[Abstract]:With the arrival of big data era, new technologies emerge in endlessly, the development of social network is in full swing.Weibo is one of the most popular social platforms, with a large group of users, generating countless hot messages every day.In Weibo, people can post original messages; users can find out who they are interested in in the system and become fans; retweets, comments and other behaviors greatly enrich the interactive experience between users.It also makes the interaction between Weibo users more diversified.However, the flood of information also makes it difficult for users to choose, and appears the phenomenon of information overload.Recommendation system is a bridge between users and projects. It can mine and capture users' preferences and actively recommend relevant content to users. It has been used in many scenarios.Collaborative filtering algorithm is one of the most classical algorithms. However, it relies heavily on the scoring data between users and items, and faces a severe problem of data sparsity.In Weibo, there is no user rating data, so we can not simply apply collaborative filtering algorithm to Weibo recommendation.Weibo's social network features provide more solutions to the recommendation problem, which can greatly improve the performance of recommendation by integrating social behavior, social trust, neighbor opinion, implicit semantic model and so on.This paper first studies the development of recommendation system and Weibo pays attention to personalized recommendation, introduces the technology and principle of collaborative filtering algorithm, and expounds the difficulties and challenges that the current algorithm is facing.By analyzing the relevant features, social atlas, user relationship and so on, the related features, social map, user relationship and so on are analyzed by Tencent Weibo data set, then the relevant terms concerned and recommended by Weibo are redefined, and the different social behaviors in Weibo are modeled.The whole process, technology platform and system environment of the system are also introduced.Aiming at the problem of Top-N recommendation, a Top-N recommendation algorithm for Weibo based on social similarity is proposed.According to Weibo's attention behavior, interactive behavior and history recommendation record, the similarity is calculated, and the nearest neighbor set is found out by the calculated similarity, and then the user is recommended.The accuracy recall rate and Fl-measurement of different similarity calculation methods are compared on Weibo data set. The parallel design of the algorithm is carried out on Hadoop platform using MapReduce to improve the efficiency of algorithm execution.Aiming at the problem of score prediction, a recommendation algorithm based on Weibo is proposed, which combines social trust and implicit semantic model.The historical recommendation records of users are modeled as scoring matrices, and social recommendations are introduced to calculate implicit trust between users through interactive behavior data between users (including @, comment and forwarding).The explicit trust between users is obtained from the direct social relationship of users, and the extended trust matrix is constructed by combining explicit trust with implicit trust, and the extended trust matrix is integrated into the SVD model.Finally, experiments on KDD Cup 2012 dataset show that the algorithm has better results on RMSE and MSE.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【相似文献】