个性化推荐系统中基于用户的协同过滤算法与系统架构的研究与优化

发布时间：2018-01-19 04:32

本文关键词： 推荐系统协同过滤系统架构 LDA主题模型栏目　出处：《山东大学》2017年硕士论文　论文类型：学位论文

【摘要】：信息技术和移动设备的成熟与发展,使网上服务与交易等业务越来越普及,大量的信息被聚集起来引发了数据的爆炸式增长。如何解决信息过载是人工智能与大数据时代面临的一个重要难题。而推荐系统是解决信息过载问题的一个非常有潜力的技术,它能够在用户没有明确目标的时候帮助他们发现感兴趣的新内容,是个性化服务研究领域的重要分支。在推荐系统的探究进程中,协同过滤算法被公认为最著名、应用最广泛的推荐算法,其核心思想是利用历史行为数据挖掘某种相似性来进行推荐,但与此同时也存在着诸如冷启动、数据稀疏、预测精确性不足等问题。本论文针对现有推荐系统的问题进行了以下三个方面的研究:1、对推荐系统架构进行优化,设置了自适应反馈机制,通过增加反馈环节,利用获取的用户反馈、上下文、以及群组信息不断调节融合参数,构成自适应闭环系统。使系统能够个性化地为用户选择最适于用户特征的推荐算法。2、在分析了多种推荐方法的基础上,重点研究推荐系统方法中的协同过滤算法,并针对基于用户的协同过滤算法进行了优化,提出了一种新的基于栏目与惩罚度的协同过滤推荐算法。其主要思想是采用一种基于栏目行为偏好与惩罚项因子相融合的方法来测量用户偏好相似性。该算法有效利用了属性层次、行为层次和评分层次信息,通过在MovieLens数据集上进行交叉验证仿真,实验结果表明,与协同过滤算法相比该算法大大提升了各项评测指标。3、通过对自然语言处理的研究,考虑将隐含狄利克雷分布(LatentDirichlet Allocation,LDA)与推荐系统结合。本论文提出了一种基于LDA特征提取的协同过滤推荐算法,通过主题模型生成相应主题以及在每个主题上的分布概率,以方便建立物品-特征向量。最后通过本论文提出的用户相似度计算方法,即加入权重距离因子的改进用户相似度计算方法,生成最近邻居集完成推荐。本论文按照数据挖掘基本流程在豆瓣图书数据集上完成实验仿真,实现了利用文本描述信息进行推荐的目标,并且获得了良好的推荐效果。通过以上研究,为推荐系统的进一步广泛应用提供了一定的参考和依据。
[Abstract]:With the maturity and development of information technology and mobile devices, online services and transactions are becoming more and more popular. A large amount of information has been gathered and triggered the explosive growth of data. How to solve the information overload is an important problem in the era of artificial intelligence and big data. Recommendation system is a very important solution to the problem of information overload. Potential technology. It can help users find new content when they have no clear goal. It is an important branch of personalized service research field. Collaborative filtering algorithm is recognized as the most famous and most widely used recommendation algorithm. Its core idea is to use historical behavior data mining some similarity to recommend, but at the same time, there are some such as cold start. In this paper, we focus on the following three aspects: 1, optimize the architecture of recommendation system and set up adaptive feedback mechanism. By adding feedback links, using the obtained user feedback, context, and group information to constantly adjust the fusion parameters. An adaptive closed-loop system is constructed, which enables the system to personalize the recommendation algorithm. 2, which is the most suitable for the user. Based on the analysis of various recommendation methods. The collaborative filtering algorithm in recommendation system is studied, and the user-based collaborative filtering algorithm is optimized. A new collaborative filtering recommendation algorithm based on column and punishment degree is proposed. The main idea of this algorithm is to measure the similarity of user preference based on the combination of column behavior preference and punishment factor. The effect takes advantage of the attribute hierarchy. Behavior level and score level information, through the MovieLens data set cross-validation simulation, experimental results show that compared with the collaborative filtering algorithm, the algorithm greatly improved the evaluation index. 3. Based on the study of natural language processing, the implicit Dirichlet Allocation is considered. In this paper, a collaborative filtering recommendation algorithm based on LDA feature extraction is proposed. The topic model is used to generate the corresponding topic and the distribution probability on each topic. In order to facilitate the establishment of object-feature vector. Finally, through the user similarity calculation method proposed in this paper, that is, adding the weight distance factor to improve the user similarity calculation method. According to the basic flow of data mining, this paper completes the experiment simulation on the Douban book data set, and realizes the goal of using the text description information to recommend. Through the above research, it provides a certain reference and basis for the further application of the recommendation system.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【相似文献】