基于权重调节的矩阵补全协同过滤算法研究

发布时间：2018-04-09 01:38

本文选题：权重调节　切入点：矩阵补全　出处：《吉林大学》2016年硕士论文

【摘要】：随着互联网技术和信息技术的蓬勃发展,个性化推荐技术的研究势如破竹,而作为个性化推荐系统中最著名,也是被研究最多的协同过滤推荐算法,现在以及将来的一段时间内,都将成为电子商务系统领域重要的研究方面。但是伴随着应用范围的扩大和应用环境的变化,加上实际网站中对推荐精度和推荐效率的严苛要求,协同过滤算法的一些缺点也显现出来。其中比较典型的有推荐实时性问题、新用户或新项目的冷启动问题、推荐结果的准确性问题等等。传统的协同过滤算法是在计算两个用户或者两个项目之间的相似性时,都是直接通过对已有的评分信息进行处理,得出用户和用户(项目和项目)之间的相似度。很少考虑评分以外的其他信息对相似度的影响。而基于协同过滤的改进算法应该尽可能多地利用系统中已有的信息来预测更加贴近目标用户兴趣的推荐结果,这样综合了多方面因素的考量可以有效地提高算法结果的准确率。本文是在协同过滤的研究基础之上,添加了针对于评分的时间因素的考量以及项目本身属性类别的考量,来有效地提高系统算法的准确度。具体工作如下:首先,通过分析考虑时间因素对协同过滤算法的影响,将用户对项目评分时的时间值融入到用户对项目的评分中,即对原数据集中用户对项目的评分值进行时间上的权值修正,提出考量时间因素对原数据集中的评分进行修正的协同过滤。其次,利用传统的协同过滤算法在计算项目的相似度的时候,仅仅考虑到用户对项目的评分信息,而没有考虑项目本身的分类属性信息对相似度的影响。因此本文将项目之间的属性差异度融入到计算项目之间的相似度公式中,并以此为影响力因子作为对目标用户预测结果的修正。最后,将改进算法应用到Movie Lens数据集上,设计实验,首先对比用余弦相似度公式和pearson相关系数公式计算项目间相似度,找到绝对误差率MAE较小的那个;然后将融合时间衰减曲线的评分修正和基于项目自身属性标签的权值调节分别与未进行改进的协同过滤算法进行推荐结果的比较,证明了算法的改进可以有效的提高系统的推荐正确率和推荐精准率。
[Abstract]:With the rapid development of Internet technology and information technology, the research of personalized recommendation technology is in full swing, and as the most famous recommendation system, it is also the most studied collaborative filtering recommendation algorithm.Now and in the future, it will become an important research area in the field of e-commerce systems.But with the expansion of application scope and the change of application environment, coupled with the harsh requirements of recommendation accuracy and recommendation efficiency in the actual website, some shortcomings of collaborative filtering algorithm also appear.Some typical problems include recommendation real-time problem, cold start-up problem of new user or new project, accuracy of recommendation result and so on.When calculating the similarity between two users or two items, the traditional collaborative filtering algorithm directly processes the existing scoring information to get the similarity between users and users (items and items).Little consideration is given to the effect of information other than scores on similarity.The improved algorithm based on collaborative filtering should make use of the existing information in the system as much as possible to predict the recommended results which are closer to the interests of the target users, which can effectively improve the accuracy of the algorithm results by synthesizing various factors.Based on the research of collaborative filtering, this paper adds the consideration of the time factor and the attribute category to improve the accuracy of the system algorithm.The specific work is as follows: firstly, by analyzing the influence of time factors on the collaborative filtering algorithm, the time value of the user scoring the item is integrated into the user's score of the item.That is to say, the user in the original data set can correct the score of the item in time, and a collaborative filtering method considering the time factor to correct the score of the original data set is put forward.Secondly the traditional collaborative filtering algorithm is used to calculate the similarity of items only taking into account the users' scoring information and not considering the impact of the items' classification attribute information on the similarity.In this paper, the attribute difference between items is incorporated into the similarity formula of the project, and it is used as the influence factor to modify the prediction results of the target users.Finally, the improved algorithm is applied to the Movie Lens data set, and the experiment is designed. Firstly, the similarity between items is calculated by using cosine similarity formula and pearson correlation coefficient formula, and the one with low absolute error rate (MAE) is found.Then the score correction of the fusion time attenuation curve and the weight adjustment based on the attribute label of the item itself are compared with the unimproved collaborative filtering algorithm.It is proved that the improved algorithm can effectively improve the recommendation accuracy and accuracy of the system.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.3

【相似文献】