基于二部图的推荐算法研究

发布时间：2018-01-12 17:44

本文关键词：基于二部图的推荐算法研究　出处：《南京航空航天大学》2015年硕士论文　论文类型：学位论文

【摘要】：电子商务的快速发展使互联网信息呈爆炸性增长,持续增加的商品和服务在丰富用户选择的同时,不可避免的引发了信息过载(Information Load)问题:用户在找到自己需要的商品之前,往往需要浏览大量的无关信息,信息使用效率下降。在此情况下,推荐系统应运而生。20世纪90年代以来,出现了大量的推荐系统方面的研究成果,本文主要研究了推荐系统领域的一类算法——基于二部图的推荐算法。复杂网络推荐算法的经典算法是基于二部图网络结构的推荐算法,但以往研究通常不考虑用户对项目的显式评分或是直接排除低分项目,造成信息丢失,数据更加稀疏。针对此问题,本文改进了这一推荐算法,在用户推荐能量流动时充分考虑用户对项目的显式评分,首次提出用户兴趣相似系数和用户评分代表能力两个概念。实验结果表明,改进后的算法能有效提高推荐算法的准确性,使推荐服务更好地满足用户的偏好需求。传统协同过滤推荐算法存在可扩展性差,多样性低,多样性与准确性不平衡问题。本文根据复杂网络理论,将显式用户行为数据与隐式用户行为数据分别映射为有权二部图与无权二部图。在网络局部范围内,首次使用选择性随机游走代替传统的Pearson相关系数、余弦相似性或修正的余弦相似性方法来计算用户之间的相似度,通过制定游走过程中的目标节点选择策略,避免了明星用户与多个用户产生弱连接,从而有效提高了算法的多样性。此外,选择性随机游走的鲁棒性保证了推荐算法的可扩展性。本文采用Movielens和Netflix显式数据集以及TMALL隐式数据集进行实验。结果表明:无论是在显式数据集上的实验还是隐式数据集的实验,与经典的基于用户的协同过滤的推荐算法和基于幂律函数的协同过滤推荐算法相比,本文算法在保持推荐准确性的同时,提高了推荐的多样性,缓解了推荐多样性与准确性不平衡的问题。
[Abstract]:The rapid development of E-commerce makes the Internet information exploding, the increasing of goods and services in the rich user choices at the same time, inevitably lead to information overload (Information Load): before the users find the goods they need, often need to browse a large number of irrelevant information, decrease the efficiency of information application. In this case, since the.20 century recommendation system came into being in 90s, the research results of many aspects of the recommendation system, this paper mainly studies the field of recommender system algorithms - recommendation algorithm based on the two plans. The classical algorithm of complex network recommendation algorithm is recommended two map algorithm based on network structure, but previous studies usually do not consider the user explicit ratings of the items or directly out of the low, caused by the loss of information, the data is more sparse. Aiming at this problem, this paper has improved the A recommendation algorithm, considering the explicit user ratings of the items in the user recommended energy flow, first proposed the user interest similarity and user rating on behalf of two concepts. The experimental results show that the improved algorithm can effectively improve the accuracy of the recommendation algorithm, make the recommendation service satisfy the user preference of traditional collaborative. Filtering algorithms are poor in scalability, low diversity, diversity and accuracy imbalance. Based on complex network theory, the explicit and implicit user behavior data type user behavior data are mapped to the right two to two map and map. In the local area network for the first time, the use of selective random walk the Pearson correlation coefficient instead of the cosine similarity or adjusted cosine similarity to calculate the similarity between users through the method, the target node in the process of making migration Selection strategy to avoid the star user and multiple users have weak connections, so as to improve the diversity of the algorithm. In addition, the robustness of the selective random walk to ensure the scalability of the recommendation algorithm. This paper uses the Movielens and Netflix data sets and TMALL explicit implicit data set. Experimental results show that: no matter in the data set on the explicit or implicit experimental data sets and experimental, and collaborative filtering recommendation algorithm and compared the collaborative filtering recommendation algorithm based on user power-law function based on the classic, the algorithm can maintain recommendation accuracy at the same time, improve the diversity of recommendations, alleviate the recommended diversity and accuracy the balance of the problem.

【学位授予单位】：南京航空航天大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP391.3

【相似文献】