协同过滤混合相似度推荐机制研究与设计

发布时间：2018-04-24 02:31

本文选题：推荐系统 + 协同过滤　；参考：《广东工业大学》2016年硕士论文

【摘要】：互联网的普及和信息技术的快速发展,网络数据呈指数级增长,尤其近年社交网站和电子商务网站的兴起与迅猛发展,出现了“信息资源过载”和“数据爆炸”的现象。这些现象已经成为商家与用户不容忽视的重要难题,而个性化推荐服务的出现是解决这一现状的有效手段。在推荐算法研究中,协同过滤算法是目前应用最为广泛的个性化推荐技术之一,如基于用户的与基于项目的协同过滤机制与算法等。然而,目前的大多数推荐系统都存在用户冷启动问题和数据稀疏性问题。此外,只采用传统的相似度来计算项目最近邻居,会忽略用户行为,未能关注用户的各种兴趣爱好。同时,计算寻找用户最近邻居时忽略用户之间的信任问题,也会对推荐质量有所影响。因此,优化相似度计算,更合理的解决数据稀疏性与冷启动问题,提高推荐准确率,为用户提供更人性化的推荐服务已成为协同过滤推荐领域的热点研究方向。本文根据以上传统协同过滤存在的相关问题,提出了一种基于协同过滤的混合相似度推荐算法。该算法考虑用户行为,采用混合模型,改进相似度度量计算,通过将项目属性相关性和修正余弦相似性进行线性组合,提出一种混合相似度计算方法来计算最近邻居项目集。同时,研究分析传统协同推荐系统没有引入用户间的信任关系,通过用户-信任关系矩阵计算用户之间的信任度,再结合上述的混合相似度来计算用户最近邻居,最终将用户的评分相似性和用户的信任度相结合,形成新的相似度度量方法来对目标项目进行预测评分,最终形成TOP-N推荐对象列表向用户推荐。论文通过算法比较,在Epinions公开数据集上将所研算法与混合相似度的用户多兴趣推荐算法、基于用户的协同过滤推荐算法、基于用户信任的协同过滤推荐算法、基于信任的用户聚类推荐算法四种算法进行了对比。实验结果表明,本文提出的基于协同过滤的混合相似度推荐算法比其他四个算法在平均绝对误差值上有一定的提高,证明了本算法的可行性与有效性。该推荐算法不仅提高了推荐的准确率,而且在一定程度上可以解决用户的冷启动问题,让用户有更人性化的推荐体验,有助于推动推荐机制的发展。
[Abstract]:With the popularization of the Internet and the rapid development of information technology, the network data is increasing exponentially. In recent years, especially the rise and rapid development of social and electronic commerce websites, the phenomenon of "information resources overload" and "data explosion" appears. These phenomena have become an important problem that can not be ignored by merchants and users, and the emergence of personalized recommendation service is an effective means to solve this situation. In the research of recommendation algorithms, collaborative filtering algorithm is one of the most widely used personalized recommendation technologies, such as user-based and project-based collaborative filtering mechanisms and algorithms. However, most recommendation systems have cold boot problem and data sparsity problem. In addition, only traditional similarity is used to calculate the nearest neighbor of the project, which ignores user behavior and fails to pay attention to various interests of users. At the same time, the problem of trust between users is ignored when searching for the nearest neighbor of users, which will affect the quality of recommendation. Therefore, optimizing similarity calculation, solving the problem of data sparsity and cold start more reasonably, improving recommendation accuracy and providing users with more humanized recommendation service have become the hot research direction in collaborative filtering recommendation field. In this paper, a hybrid similarity recommendation algorithm based on collaborative filtering is proposed. Considering user behavior, the algorithm uses a hybrid model to improve similarity measurement. By combining item attribute correlation with modified cosine similarity, a hybrid similarity calculation method is proposed to calculate the nearest neighbor itemsets. At the same time, the paper analyzes that the traditional collaborative recommendation system does not introduce the trust relationship between users. The trust degree between users is calculated by the user-trust matrix, and the user nearest neighbor is calculated by combining the hybrid similarity mentioned above. Finally, a new similarity measure method is formed to predict the target items by combining the user's rating similarity and user's trust, and finally the list of TOP-N recommendation objects is formed to recommend to the user. Through the comparison of algorithms, this paper compares the proposed algorithm with mixed similarity recommendation algorithm, collaborative filtering recommendation algorithm based on user, collaborative filtering recommendation algorithm based on user trust, and collaborative filtering recommendation algorithm based on user trust. The trust-based user clustering recommendation algorithm is compared. The experimental results show that the proposed hybrid similarity recommendation algorithm based on collaborative filtering is better than the other four algorithms in the mean absolute error value, which proves the feasibility and effectiveness of this algorithm. This recommendation algorithm not only improves the accuracy of recommendation, but also can solve the cold start problem of users to a certain extent, make users have more humanized recommendation experience, and help to promote the development of recommendation mechanism.
【学位授予单位】：广东工业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.3

【相似文献】