《一种基于联合聚类的协同过滤推荐算法研究》由会员分享，可在线阅读，更多相关《一种基于联合聚类的协同过滤推荐算法研究（51页珍藏版）》请在金锄头文库上搜索。
1、 I 华 中 科 技 大 学 硕 士 学 位 论 文 摘 要 摘 要 随着互联网中信息量的急剧增长，人们在海量信息中快速获得自己感兴趣的信息变得越来越困难。如电子商务、社交媒体、音乐、视频、问答等站点，信息量越来越庞大，而用户所关心与关注的往往是很少的一部分。个性化技术强调从用户兴趣出发，针对不同用户提供特定的个性化服务。其中，推荐系统根据用户历史信息发现用户的兴趣，从而主动向用户推荐其可能感兴趣的资源，帮助用户快速获取有用信息。 协同过滤技术是一种基于用户显性评分行为的推荐技术，是应用最早且最为成功的推荐技术之一。然而，随着系统中用户数量与项目数量的与日俱增，其面临着两个主要挑战： （1）提
2、高协同过滤推荐算法的扩展性，实时处理大量动态数据； （2）降低系统中评分数据的稀疏性，提高推荐精度。 研究了一个改进的协同过滤算法同时基于用户与项目两个维度进行联合聚类的协同过滤方法。聚类后的矩阵规模远远小于原始评分矩阵，且同一类别内部的评分具有相似的模式，可快速灵活的进行评分预测。针对该离线聚类模型的增量更新机制提出了改进方法，将新的评分数据加入平均评分最接近的类别。针对基于加权非负矩阵分解的评分预测方法中初始值填充问题，提出了结合基线估计的初值填充方法。最后在 MovieLens 数据集上进行聚类效果实验与评分预测实验，验证了该联合聚类方法的有效性与实用性。 关键词关键词：推荐系统，协同过
3、滤，联合聚类，加权非负矩阵分解 II 华 中 科 技 大 学 硕 士 学 位 论 文 AbstractAbstract With the rapid growth of the information on Internet, people find it increasingly difficult to quickly get the information they are desiring. Such as e-commerce, social media, music, video, quizzes and other sites, the amount of information
4、 is growing, but a single users interest and concern is often a small part. Personalization technologies focus on different interest of different users, and provide specific personalized service for different users. Among them, the recommender system based on users history information to find out us
5、ers interest in order to make appropriate recommendations for users according to their personal interest and help users quickly access useful information. Collaborative filtering techniques based on the most similar users interests to make recommendations, is the first and the most successful techno
6、logies applied in recommender systems. However, with the increasing amount of users and items, it faces two major challenges: (1) to find new collaborative filtering algorithms to improve scalability and process large scale and dynamic datasets in a real-time environment ; (2) to reduce sparsity in
7、recommender systems , and improve the quality of recommendations. An improved collaborative filtering algorithm - co-clustering based on two dimensions: users and items was researched. On one hand, Matrix after co-clustering is much smaller than the size of the original score matrix; on the other ha
8、nd, users scores within the same class have similar patterns and predicting is fast and flexible. The co-clustering model also proposes an incremental updating method to improve real-time dynamic data processing, which takes on within the internal of corresponding classes. Weighted non-negative matr
9、ix factorization on the small scale clusters after co-clustering, has a good speed and satisfied accuracy. Experiments on the MovieLens datasets compared the performance of the new algorithm and traditional clustering methods, as to verify the validity of this method and its values in large-scale re
10、al-time recommendation systems. Keywords：Recommender System, Collaborative Filtering, Co-clustering, WNMF III 华 中 科 技 大 学 硕 士 学 位 论 文 目目 录录 摘 要 . I Abstract . II 1 绪论 1.1 研究背景 . (1) 1.2 目的与意义 . (1) 1.3 国内外研究概况 . (2) 1.4 本文研究内容 . (5) 2 推荐系统与协同过滤技术 2.1 推荐系统简介 . (6) 2.2 传统的协同过滤技术 . (7) 2.3 基于聚类的协同过滤技术 . (13) 2.4 本章小结 . (18) 3 基于联合聚类的协同过滤算法 3.1 联合聚类的概念 . (19) 3.2 基于联合聚类的协同过滤 . (20) 3.3 基于加权非负矩阵分解的评分预测 . (29) 3.4 本章小结 . (32) 4 实验与分析 4.1 实验数据集 . (33) 4.2 聚类效果实验 . (34) 4.3 评分预测实验 . (36) 4.4 本章小结 .
