基于社交网络的上下文感知推荐算法

发布时间：2018-11-17 13:34

【摘要】：随着信息技术的快速发展,人类社会已经由信息贫乏的时代进入了信息过载时代。面对互联网上的海量信息,一方面用户很难从中找到自己真正感兴趣的信息,另一方面,信息的生产者也很难找到对其真正感兴趣的用户,从而使自己的信息受到关注。推荐系统通过分析用户行为数据,提取出用户偏好,给用户提供个性化的推荐内容,在很多的网络应用(比如电子商务网站亚马逊,淘宝网以及社交网站Linked,Facebook,人人网等)中,已经成为了一个很有前途的处理信息过载的工具。目前,推荐系统研究领域应用较多的推荐算法包括基于用户的协同过滤推荐算法、基于物品的协同过滤推荐算法、基于隐语义模型的推荐算法、基于上下文信息的推荐算法以及基于社交网络的推荐算法。其中应用最为广泛的是协同过滤(CF)推荐,它通过挖掘相似用户或项目的历史行为数据来预测目标用户的偏好。尽管协同过滤推荐算法已经在业界得到了广泛应用,但传统的协同过滤技术只利用了“用户-项目”二元关系而未考虑其它信息。当信息规模越来越大时,它的性能就遇到了很大挑战,比如数据的稀疏性(即缺乏足够数量的相似用户或项目),由数据稀疏性及信息源的同质化造成的推荐质量下降。本文主要研究上下文感知推荐算法,对上下文的概念,上下文感知推荐系统的研究现状,社交网络数据及用户行为数据进行了详细介绍。重点研究了上下文信息的提取及对多种上下文信息的处理,对社交网络数据的处理及用户相似度的计算,并提出了基于上下文提取的感知推荐算法以及在此基础上引入社交网络数据的基于社交网络的上下文感知推荐算法。实际应用中存在着多种类型的上下文信息,但并不是每种上下文信息对于用户的偏好都能起到同样的影响。基于上下文提取的感知推荐算法通过比较传统推荐模型在不同上下文片段上的性能来识别出那些能影响用户偏好的上下文片段,应用随机决策树算法将含有不同类型上下文信息的评分进行分割,所产生的子矩阵中的评分处于相似的上下文中,彼此之间相关度更高。在树的叶子结点应用矩阵分解,通过求解目标函数来预测目标用户对项目的评分。社交网络信息是另一类能够对用户偏好产生重要影响的信息。基于社交网络的上下文感知推荐算法引入了一个社交正则化项,通过学习用户好友的偏好来预测用户的偏好。为了识别有着相似偏好的好友,提出一种融入上下文信息的皮尔森相关系数(pcc)来度量用户相似度。理论分析与实验结果表明基于上下文提取的感知推荐算法及基于社交网络的上下文感知推荐算法在准确率上较传统的推荐算法在性能方面有明显的提高。
[Abstract]:With the rapid development of information technology, human society has entered the age of information overload from the era of poor information. In the face of the mass of information on the Internet, it is difficult for users to find the information they are interested in. On the other hand, it is difficult for the producers of information to find the users who are interested in it. By analyzing user behavior data, the recommendation system extracts user preferences, provides personalized recommendation content to users, and works in many web applications (e. G. Amazon, Taobao and social networking site Linked,Facebook,) Renren, etc., has become a promising tool for handling information overload. At present, many recommendation algorithms are used in the research field of recommendation system, including user-based collaborative filtering recommendation algorithm, object-based collaborative filtering recommendation algorithm, and recommendation algorithm based on hidden semantic model. Recommendation algorithm based on context information and recommendation algorithm based on social network. The most widely used (CF) recommendation is collaborative filtering, which predicts the preferences of target users by mining historical behavior data of similar users or projects. Although collaborative filtering recommendation algorithm has been widely used in the industry, the traditional collaborative filtering technology only uses the "user-item" binary relationship without considering other information. When the scale of information becomes larger and larger, its performance meets great challenges, such as the sparsity of data (that is, the lack of a sufficient number of similar users or projects), and the deterioration of recommendation quality caused by the sparsity of data and the homogeneity of information sources. This paper mainly studies the context-aware recommendation algorithm, introduces the concept of context, the research status of context-aware recommendation system, social network data and user behavior data in detail. It focuses on the extraction of context information and the processing of various context information, the processing of social network data and the calculation of user similarity. A context-aware recommendation algorithm based on context extraction and a context-aware recommendation algorithm based on social network data are proposed. There are many types of context information in practical applications, but not every context information has the same effect on user preferences. Context-based aware recommendation algorithms identify those contexts that affect user preferences by comparing the performance of traditional recommendation models on different context segments. The random decision tree algorithm is used to segment the score with different types of context information. The score in the generated submatrix is in the same context and the correlation between each other is higher. The matrix decomposition is applied to the leaf node of the tree to predict the target user's score of the item by solving the objective function. Social network information is another type of information that can have an important impact on user preferences. The context-aware recommendation algorithm based on social networks introduces a social regularization item to predict the preferences of users by learning the preferences of their friends. In order to identify friends with similar preferences, a Pearson correlation coefficient (pcc) is proposed to measure user similarity. Theoretical analysis and experimental results show that the performance of context-based perceptive recommendation algorithm and context-aware recommendation algorithm based on social network is significantly higher than that of traditional recommendation algorithm.
【学位授予单位】：沈阳建筑大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.3

【参考文献】