基于用户兴趣和领域最近邻的混合推荐算法研究

发布时间：2018-03-23 09:27

本文选题：协同过滤　切入点：用户兴趣　出处：《安徽理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：面对大数据的时代,怎么从杂乱无章的信息海洋里准确的推荐给用户感兴趣的信息,这将是推荐算法研究的主要任务。最为经典的两个推荐算法是基于内容过滤和协同过滤推荐算法,但再经典的推荐算法也有自己的缺点。数据稀疏性和冷启动是协同过滤推荐算法的主要问题。基于内容过滤的推荐算法有一个比较严重的问题,那就是新用户问题,因为该算法并未考虑用户的兴趣改变对推荐效果的影响。当系统中新增一个用户时,新增用户的历史浏览记录是不存在的,它将无法对新增用户做出正确的推荐。针对这些,本文提出一种结合了用户兴趣和领域最近邻的的混合推荐算法(UIDNN),用于个性化服务推荐。首先,考虑用户的兴趣偏好不是永远不变的。用户的兴趣偏好随着时间的变化跟人类对于基本事物的遗忘规律很类似。引入非线性逐步遗忘函数求取用户对商品项目的兴趣度。然后根据用户-商品属性标签集合形成用户-兴趣度集合,对用户-商品项目评分集合中未评价商品项目采用平均值法进行填充、已评价商品项目进行互补形成用户-兴趣度矩阵,降低了数据的稀疏性。其次,引入"属性领域最近邻"方法查找目标用户的最近邻,在查找最近邻居时,根据用户-兴趣度集合去降低算法的在线计算量。这种做法主要是通过判断目标用户的邻居有没有这个推荐能力,从而不去考虑那些对目标用户无推荐能力的用户。预测未评价商品评分,采用用户-兴趣度集合的余弦相似度计算用户的相似度;最后把与目标用户相似度大小在前N位的项目推荐给目标用户。基于这些对目标用户进行推荐。通过实验,本文提出的基于用户兴趣和领域最近邻的混合推荐算法(UIDNN)跟相似度计算方法为皮尔逊相似度(Pearson)、余弦相似度(cos)两种传统的基于用户的协同过滤推荐算法进行比较平均绝对误差(MAE),由实验结果图可以看出,本文提出的基于用户兴趣和领域最近邻的混合推荐算法(UIDNN)有较小的MAE,说明本文提出的UIDNN算法有较高的推荐质量。
[Abstract]:In the face of big data's time, how to accurately recommend information of interest to users from a messy ocean of information, This will be the main task in the research of recommendation algorithms. The two most classical recommendation algorithms are based on content filtering and collaborative filtering recommendation algorithms. However, the classical recommendation algorithm also has its own shortcomings. Data sparsity and cold start are the main problems of collaborative filtering recommendation algorithm. There is a serious problem in the content filtering recommendation algorithm, that is, the problem of new users. Because the algorithm does not take into account the influence of the user's interest change on the recommendation effect. When a new user is added to the system, the historical browsing record of the new user does not exist, and it will not be able to make the correct recommendation to the new user. In this paper, we propose a hybrid recommendation algorithm, which combines the interests of users and the nearest neighbor of the domain, for personalized service recommendation. The change of user's interest preference over time is very similar to the law of human's forgetting of basic things. The nonlinear stepwise forgetting function is introduced to find the user's item of merchandise. Interest. Then form a user-interest set based on the user-commodity attribute label set, The average value method is used to fill the unevaluated items in the user-commodity item score set. The evaluated commodity items complement each other to form the user-interest matrix, which reduces the sparsity of the data. The nearest neighbor of the property domain method is introduced to find the nearest neighbor of the target user. Based on the user-interest set to reduce the online computation of the algorithm. This approach is mainly by judging whether the neighbor of the target user has the ability to recommend. Therefore, the users who have no recommendation ability to the target users are not considered. The users' similarity is calculated by using the cosine similarity of the user-interest set. Finally, we recommend the items with the first N bit similarity to the target users. Based on these, we recommend the target users. This paper proposes a hybrid recommendation algorithm based on user interest and domain nearest neighbor (UIDNN) and its similarity calculation methods are Pearsonian (cosine similarity) and Pearsonian (Pearsonian), two traditional user-based collaborative filtering and recommendation algorithms are compared. The mean absolute error can be seen from the diagram of the experimental results. The proposed hybrid recommendation algorithm based on user interest and domain nearest neighbor has a small mae, which shows that the proposed UIDNN algorithm has high recommendation quality.
【学位授予单位】：安徽理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【参考文献】