基于两阶段行为模式的协同过滤推荐算法研究

发布时间：2018-04-02 01:28

本文选题：协同过滤　切入点：Top-N推荐　出处：《扬州大学》2017年硕士论文

【摘要】：由于互联网中信息爆炸式的增长,导致用户很难直接发现有用的信息。为此,根据用户的历史行为数据进行建模的推荐系统吸引了学者的广泛关注,可以一定程度上缓解信息过载的问题。各种类型的推荐算法中,协同过滤推荐算法以其可理解性强、无需语义分析等特点成为目前使用最广泛的推荐算法之一。但是,随着数据规模的增大,协同过滤算法面临着数据稀疏,实时性以及准确性和多样性的权衡等方面的问题。另外,与之前的评分预测相比,Top-N推荐形式更符合当前协同过滤推荐算法的需求。因此,本文主要针对协同过滤算法存在的部分问题,研究协同过滤算法在Top-N的推荐算法。本文主要的研究工作是:(1)提出了基于两步预测的二分网络Top-N推荐算法。已有的基于二分网络的协同过滤推荐算法只考虑了用户选择,忽略了用户评分。利用本文的用户行为中存在的两阶段的行为,提出了基于两步预测的二分网络Top-N推荐算法。该算法先利用NBI算法预测用户对产品进行评分的概率,然后利用两步预测将其与协同过滤算法结合进行推荐。在MoiveLens数据集上的实验表明,该算法提高了推荐的准确度。(2)提出了一种基于属性比重相似性的两步预测Top-N推荐算法。在处理极端不均匀和稀疏的用户评分数据时,传统的协同过滤推荐算法不能很好的进行相似性计算。为此,考虑用户对某一类的产品更感兴趣和一个产品可能同属于不同的属性,同时,结合用户两阶段行为模式,本文提出一种基于属性比重相似性的两步预测Top-N推荐算法。在MoiveLens数据集上的实验表明,算法能提高协同过滤算法的Top-N推荐的准确率和多样性。(3)基于Spark框架的属性比重相似性的两步预测推荐算法的并行化实现。结合Spark框架在实现并行化方面的优势和协同过滤推荐算法的实时性,本文实现了基于Spark框架的属性比重相似性的两步预测推荐算法的并行化,达到通过架构的方式提高推荐算法实效性的目标。在MoiveLens数据集上的实验结果表明,该算法能够提高运算的速率。
[Abstract]:Due to the explosive growth of Internet information, it is difficult to directly cause the user to find useful information. Therefore, according to the history data of user recommendation system modeling has attracted wide attention of scholars, can to some extent alleviate the problem of information overload. Various types of recommendation algorithm, collaborative filtering algorithm which can understand strong, no characteristic of semantic analysis has become one of the most widely used recommendation algorithm. However, with the increasing size of the data, the collaborative filtering algorithm is faced with sparse data, real-time and accuracy and diversity trade-off problems. In addition, with the previous score compared to Top-N recommendation form conforms to the current collaborative filtering recommendation algorithm needs. Therefore, this paper focuses on the problems existing collaborative filtering algorithm, research on collaborative filtering algorithm in recommendation algorithm Top-N the. The main research work is: (1) proposed the two network Top-N two step prediction algorithm based on collaborative filtering algorithm. The existing two network only considers the user selection based on ignoring the user score. Using the two stage of the user behavior in the line, put forward two network Top-N two step prediction based algorithm. The algorithm first uses NBI algorithm to predict the user scoring probability of the product, and then use the two step forecast the recommendation combined with collaborative filtering algorithm. Show the experiment on MoiveLens data set, this algorithm can improve the accuracy of the recommendation. (2) proposed a a similar proportion of the two step prediction based on attribute Top-N recommendation algorithm. In dealing with extreme uneven and sparse user rating data, the traditional collaborative filtering algorithm is not very good for similarity calculation. Therefore, consider the use of Users of a particular type of product and more interested in a product may belong to different attributes, at the same time, the two stage combined with user behavior model, this paper presents a similarity based on the attributes of the proportion of two step prediction Top-N recommendation algorithm. Show the experiment on MoiveLens data set, the algorithm can improve the accuracy of collaborative filtering algorithm the Top-N recommendation and diversity. (3) the parallel two step prediction attributes of Spark frame similarity recommendation algorithm based implementation. Combining with the Spark framework in the real-time realization of parallel advantages and collaborative filtering algorithm, this paper realized the parallelization of similar two step predictive recommendation algorithm of attribute the proportion based on the Spark framework, to improve the effectiveness of the algorithm recommended by the target through the architecture. In the MoiveLens data set. The experimental results show that the algorithm can improve the computing speed.

【学位授予单位】：扬州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【参考文献】