基于评分选取技术的推荐算法研究

发布时间：2018-11-01 20:40

【摘要】：推荐系统已经成为大数据时代最重要的信息过滤工具之一,它可以帮助用户从海量数据中迅速定位有价值的信息,并以用户可能感兴趣的物品列表的方式推荐给用户。互联网爆炸式的信息量以及用户和物品数量的快速增长使推荐系统面临着诸多挑战,可扩展性便是其中的主要挑战之一。协同过滤是推荐系统领域应用最成功、最广泛的技术。目前,很多学者为了提升协同过滤算法的可扩展性,提出了多种基于聚类和基于并行技术的方案。通常,他们会在推荐算法的建模阶段使用全部的用户评分数据,而没有考虑这些数据的质量因素,而且已有的论文大都是针对基于近邻的协同过滤算法的可扩展性。本文从输入源数据集的角度出发,提出观点:并不是所有的用户行为数据都对最终的预测模型作出了同样的贡献,尤其是对那些拥有大量行为的活跃用户而言。本文认为,对于活跃用户,部分具备代表性的行为数据已经可以包含足够的信息来对用户作出准确的建模,在更短的时间内得到一个好的推荐结果。基于上述观点,本文首先通过一系列的实验探索了推荐算法建模阶段使用的用户行为数量和推荐算法性能之间的关系,提出了基于评分选取的推荐算法。特别地,本文的所有实验均同时考虑了评分预测和TopN推荐任务。随后,本文提出了一个综合考虑用户和电影两方面因素的通用评分选取框架,并且提出了基于划分的3种评分选取策略和基于统计学与信息论的5种评分选取策略,来为每一个用户选取其最具代表性的评分。最后,本文在MovieLens和Netflix数据集上做了大量的实验,实验结果表明仅使用活跃用户的部分代表性行为可以在达到预期推荐精度的同时降低算法的运行时间,由此提升了推荐系统的可扩展性,而且本文提出的方案适用于所有的协同过滤算法。
[Abstract]:Recommendation system has become one of the most important information filtering tools in big data era. It can help users quickly locate valuable information from massive data and recommend it to users in the form of lists of items that may be of interest to users. The explosive amount of information on the Internet and the rapid growth of the number of users and items make the recommendation system face many challenges, of which scalability is one of the main challenges. Collaborative filtering is the most successful and widely used technology in the field of recommendation system. At present, in order to improve the scalability of collaborative filtering algorithms, many scholars have proposed many schemes based on clustering and parallel technology. Usually, they use all the user rating data in the modeling phase of the recommendation algorithm, without considering the quality factors of the data, and most of the existing papers focus on the scalability of the collaborative filtering algorithm based on the nearest neighbor. From the point of view of input source dataset, this paper puts forward the point of view: not all user behavior data make the same contribution to the final prediction model, especially for those active users who have a large number of behaviors. This paper holds that for active users, some representative behavioral data can already contain enough information to model users accurately and get a good recommended result in a shorter time. Based on the above viewpoint, this paper first explores the relationship between the number of user behaviors and the performance of the recommendation algorithm in the modeling phase of the recommendation algorithm through a series of experiments, and proposes a recommendation algorithm based on the selection of the score. In particular, all experiments in this paper considered both score prediction and TopN recommendation tasks. Then, this paper proposes a general scoring selection framework that considers both user and movie factors, and proposes three scoring selection strategies based on division and five scoring selection strategies based on statistics and information theory. To select the most representative score for each user. Finally, a large number of experiments have been done on MovieLens and Netflix datasets. The experimental results show that only using some representative behaviors of active users can reduce the running time of the algorithm while achieving the expected recommendation accuracy. This improves the scalability of the recommendation system, and the proposed scheme is suitable for all collaborative filtering algorithms.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.3

【相似文献】