基于SVD与SVM混合推荐的电影推荐系统的研究

发布时间：2018-01-26 14:09

本文关键词： 奇异值分解支持向量机 K近邻协同过滤推荐系统　出处：《太原理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着互联网2.0时代的到来,用户的各类网络信息数据与日俱增,信息过载的问题日益严重。对于单个用户而言,从纷繁复杂的网络世界中快速捕捉到自己需要的信息越来越难;对于产品提供方而言,如何集成所有用户的信息并迅速地挖掘到用户的个人潜在需求,把用户可能感兴趣的产品及时推送给用户成为大数据时代下精准营销的一大技术难题。个性化推荐技术作为解决信息过载的有效手段和重要工具应运而生,在电子商务领域及各类社交媒体平台展现出了良好的应用前景。其中,协同过滤推荐技术作为应用最早也最广泛的个性化推荐技术之一,在实际应用中取得了巨大的成功,但仍然面临着数据稀疏与冷启动,可扩展性差等制约推荐精度的严峻的问题。个性化推荐发展到现在,已经有大量优秀的专家学者提出了很多不同的算法模型来解决传统协同过滤的这些缺陷,其中混合推荐算法因其能够有效缓解传统协同过滤推荐手段单一,推荐效率不高等缺陷而成为推荐算法研究领域的热门方向,受到了越来越多的关注。本文提出的基于奇异值分解与支持向量机的混合推荐算法对传统协同过滤算法进行了一些相应的改进,主要工作如下:1.针对推荐系统中用户-项目评分数据的稀疏性问题,提出采用矩阵分解技术降维来最大化提取有效信息,分解得到三个稠密的包含用户对项目偏好信息的奇异矩阵,有效地缓解了原始评分矩阵的极端稀疏情况;2.针对推荐系统中用户及项目数量急剧增长引发的可扩展性差的问题,利用奇异值分解技术抽取用户-项目数据的关键特征,降低用户或项目的奇异向量维数,相比传统协同过滤一定程度上降低了相似度矩阵的计算复杂度,较好地解决了可扩展性差的问题;3.为了避免推荐系统用户及项目数量庞大导致的内存损耗问题,提出基于SVD及SVM的混合推荐算法,只需存储奇异值分解后的用户或者项目的奇异矩阵,用户或项目的特征向量维数大大降低,保证了推荐精确度的同时,节省了更多存储空间,这对于拥有浩如烟海数据的推荐系统无疑具有十分重大的意义;4.在Movie Lens数据集上进行的实证表明,本文提出的基于奇异值分解和SVM的混合推荐算法确实一定程度上缓解了数据稀疏,可扩展性差及推荐精度不高的问题。
[Abstract]:With the arrival of the Internet 2.0 era, users of all kinds of network information data is increasing, the problem of information overload is becoming more and more serious. It is more and more difficult to quickly capture the information we need from the complicated network world. For product providers, how to integrate the information of all users and quickly tap into the potential needs of users. Pushing products of interest to users in time has become a major technical problem of precision marketing in big data era. Personalized recommendation technology as an effective means to solve information overload and an important tool emerged as the times require. Collaborative filtering recommendation technology is one of the earliest and most widely used personalized recommendation technologies in the field of electronic commerce and various social media platforms. Great success has been achieved in the practical application, but still facing the data sparse and cold start, poor scalability and other severe problems restricting the accuracy of the recommendation. Personalized recommendation has developed to the present. A large number of excellent experts and scholars have proposed a lot of different algorithm models to solve these shortcomings of traditional collaborative filtering, among which the hybrid recommendation algorithm can effectively alleviate the traditional collaborative filtering recommendation means single. Recommendation efficiency and other shortcomings have become a hot research area in the field of recommendation algorithm. The hybrid recommendation algorithm based on singular value decomposition (SVD) and support vector machine (SVM) is proposed to improve the traditional collaborative filtering algorithm. The main work is as follows: 1. Aiming at the sparsity of user-item scoring data in recommendation system, a matrix decomposition technique is proposed to maximize the extraction of effective information. Three dense singular matrices containing user preference information are obtained, which can effectively reduce the extreme sparsity of the original scoring matrix. 2. Aiming at the problem of poor scalability caused by the rapid growth of users and projects in recommendation systems, singular value decomposition (SVD) is used to extract the key features of user-project data. To reduce the singular vector dimension of users or items, compared with traditional collaborative filtering, the computational complexity of similarity matrix is reduced to a certain extent, and the problem of poor scalability is solved. 3. A hybrid recommendation algorithm based on SVD and SVM is proposed to avoid the memory loss caused by the large number of users and items in the recommendation system. It only needs to store singular matrix of user or item after singular value decomposition. The dimension of eigenvector of user or item is greatly reduced, which ensures the accuracy of recommendation and saves more storage space. This is undoubtedly of great significance to the recommendation system with vast data; 4. The empirical results on the Movie Lens dataset show that the proposed hybrid recommendation algorithm based on singular value decomposition and SVM can alleviate the data sparsity to some extent. Poor scalability and recommendation accuracy is not high problems.
【学位授予单位】：太原理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：J943

【参考文献】