新型协同过滤推荐算法研究

发布时间：2018-02-28 15:34

本文关键词： 推荐算法协同过滤项目相似度学习社交网络标签　出处：《安徽大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着互联网的发展,推荐算法已经应用到很多领域,协同过滤推荐算法是经典的、应用广泛的推荐算法。然而传统的协同过滤推荐算法面临着很多问题,其中最严重的是冷启动问题、数据稀疏问题和扩展性问题。本文针对这些问题,对传统的协同过滤推荐算法做了一定的改进。首先,针对数据稀疏性问题,本文提出了一种基于项目相似度学习的协同过滤推荐算法。该算法首先根据项目属性相似性度量方法计算出所有项目的相似度矩阵,然后选取目标项目的前K个最相似的项目作为其初始邻近集;再将训练集中目标项目的评分向量作为期望输出,目标项目的K个邻近项目的评分向量输入到RBF神经网络中进行学习,得到项目相似度训练模型;再将测试数据集中的目标项目的K个邻近项目的评分向量输入训练模型,最后输出目标项目的预测评分向量。针对新项目冷启动问题,我们计算出新加入项目与其他项目的属性相似度,然后取出前K个最相似的项目构成邻近集并且计算出新加入项目的预测评分向量。最后取出对目标项目评分大于等于3且分数排在前N位的用户,并将目标项目推荐给这些用户。其次,针对数据稀疏性和扩展性问题,本文提出了一种基于社交网络和标签的协同过滤推荐算法。该算法将目标用户与他的朋友之间的信任度、熟悉度和标签信息反映的兴趣偏好相似度结合起来,计算出与他相似度较高的K个朋友作为邻居集合,从而为目标用户推荐喜欢的项目;然后,针对新用户冷启动问题,提出了基于朴素贝叶斯算法的模型。它利用朴素贝叶斯算法对训练集中的用户进行分类,将新用户划分到所属的类别,即求出新用户最喜欢的项目类型,然后在这种类型的项目里选择评分最高的N个项目推荐给该用户。最后,在Movielens数据集上实现基于项目相似度学习的协同过滤推荐算法,交叉实验表明,该算法在处理稀疏数据时表现出了较好的性能,并且得到了更准确的推荐结果;在Last.fm数据集上实现基于社交网络和标签的协同过滤推荐算法,与传统的算法和一些经典的算法相比,该算法具有较好的准确性和高效性。最后,在Movielens数据集上验证了项目冷启动和用户冷启动问题,实验表明算法在一定程度上解决了冷启动问题。
[Abstract]:With the development of the Internet, recommendation algorithms have been applied to many fields. Collaborative filtering recommendation algorithms are classic and widely used. However, the traditional collaborative filtering recommendation algorithms face many problems. The most serious problems are cold start problem, data sparse problem and expansibility problem. In this paper, some improvements are made to the traditional collaborative filtering recommendation algorithm. In this paper, a collaborative filtering recommendation algorithm based on item similarity learning is proposed. Then the first K similar items of the target item are selected as its initial adjacent set, and the score vector of the target item in the training set is taken as the expected output. The score vectors of K adjacent items of the target items are input into the RBF neural network for learning, and the item similarity training model is obtained, and then the score vectors of K adjacent items in the test data set are input into the training model. Finally, we output the prediction score vector of the target item. For the cold start problem of the new project, we calculate the attribute similarity between the new item and other items. Then take out the first K most similar items to form the adjacent set and calculate the predicted score vector for the new item. Finally, take out the user whose target item score is greater than or equal to 3 and scores in the top N position. Secondly, a collaborative filtering recommendation algorithm based on social networks and tags is proposed to solve the problem of data sparsity and scalability, which brings forward the trust between the target user and his friend. By combining familiarity with interest preference similarity reflected by label information, K friends with high similarity are calculated as neighbors to recommend favorite items for target users. This paper presents a model based on naive Bayes algorithm, which classifies users in training set and divides new users into categories, that is, to find out the type of items that new users like most. Finally, a collaborative filtering recommendation algorithm based on item similarity learning is implemented on the Movielens dataset. The algorithm shows better performance in dealing with sparse data and gets more accurate recommendation results, and implements collaborative filtering recommendation algorithm based on social networks and tags on Last.fm datasets. Compared with the traditional algorithm and some classical algorithms, this algorithm has better accuracy and high efficiency. Finally, the cold start problem of the project and the cold start problem of the user are verified on the Movielens data set. Experiments show that the algorithm solves the cold start problem to some extent.
【学位授予单位】：安徽大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【参考文献】