基于上下文属性信息的个性化推荐系统研究

发布时间：2018-06-12 02:40

本文选题：上下文 + 张量分解　；参考：《山东师范大学》2017年硕士论文

【摘要】：随着网络信息资源的急速增长,用户快速且准确地获取所需信息变得十分困难。搜索引擎的出现解决了用户一部分查询的困难,但是目前该工具实现不了根据用户的需求进行推荐的功能。个性化推荐系统是以用户的需求为标准。比如,商家可以根据海量的数据挖掘用户的偏好信息,将潜在客户挖掘出来进而将销售范围进一步扩大,从而拥有更多的消费群体。个性化推荐系统就是根据不同用户具有不同兴趣点这一个客观现象,对用户进行个性化推荐,使用户能够在大量信息中快速选定自己需要的商品,从而在选择商品的过程中减少不必要的挑选时间。所以,个性化推荐系统无论针对用户还是商家而言都具有实用性和价值性。协同过滤推荐算法是对用户的行为信息进行分析,将与目标用户行为信息相近的用户查找出来,依据相近用户对某些物品的偏好度去衡量目标用户对物品的偏好度,将目标用户对物品的偏好度按照从高到低进行排序,最后将结果反馈给目标用户。基于内容的推荐算法实现的主要原理是:根据对用户的特征和项目的特征的有效分析进行推荐。常用的项目特征分析建立方法包括:贝叶斯模型,神经网络模型和空间向量模型。用户的特征则是根据用户偏好的项目信息分析得出的。推荐算法可以有效的提高用户从浏览者身份到购买者身份的转化率,从而提升了销售能力。尽管这些常用的推荐算法已经取得了很大成果,但是仅仅根据单一的评分数据来挖掘相似的用户和物品,得出的推荐效果并不是很理想。现在很多学者在个性化推荐算法中加入了一些上下文属性信息,比如标签、地点等,用这些上下文属性信息来改善个性化推荐的效果。本文在阅读大量文献的基础上,对推荐算法的关键技术进行了研究,根据已有技术进行了创新型改进,并通过仿真模拟实验证明了该方案的可行性和优势性。本文的具体成果如下:(1)将用户之间共同评价的项目上下文信息和共同评价过项目的用户上下文信息融合到推荐算法当中,有效提高了推荐效果的准确率;(2)提出一种基于上下文感知和张量分解的个性化推荐算法(CATD),并在Movie lens大规模真实数据集上进行了仿真实验,验证了该算法的有效性。(3)利用核密度估计技术以及用户、项目的上下文属性信息,分别构建用户和项目的偏好模型,在偏好模型基础上提出了新的相似度计算方法,再将相似性度量值高的近邻进行融合;最后结合一定的推荐方法进行用户和项目间的推荐。
[Abstract]:With the rapid growth of network information resources, it is very difficult for users to obtain the required information quickly and accurately. The appearance of search engine solves the difficulty of querying part of the user, but at present, the tool can not realize the function of recommending according to the user's demand. Personalized recommendation system is based on the needs of users. For example, businesses can mine user preferences based on massive data, mining out potential customers, and further expand the range of sales, so as to have more consumer groups. Personalized recommendation system is based on the objective phenomenon that different users have different points of interest, so that users can quickly choose the products they need in a large amount of information. As a result, in the selection of goods in the process of reducing unnecessary selection time. Therefore, personalized recommendation system is practical and valuable for both users and merchants. Collaborative filtering recommendation algorithm is to analyze the user's behavior information, find out the users who are close to the target user's behavior information, and measure the target user's preference for some items according to the similar user's preference for some items. The target user's preference for the item is sorted from high to low, and the result is fed back to the target user. The main principle of the implementation of content-based recommendation algorithm is to make recommendations based on the effective analysis of the features of the users and the features of the items. The commonly used project feature analysis methods include Bayesian model, neural network model and spatial vector model. The characteristics of the user are analyzed according to the item information of the user preference. The recommendation algorithm can effectively improve the conversion rate from the user's identity to the buyer's identity, thus enhancing the sales ability. Although these commonly used recommendation algorithms have made great achievements, but only based on a single score data to mine similar users and items, the recommended results are not very good. Nowadays, many scholars have added some contextual attribute information, such as label, location, etc., to improve the effect of personalized recommendation. On the basis of reading a large number of literatures, this paper studies the key technologies of the recommendation algorithm, and makes an innovative improvement according to the existing technology. The feasibility and superiority of the scheme are proved by simulation experiments. The concrete results of this paper are as follows: (1) the project context information which is evaluated jointly by users and the user context information that has been evaluated jointly is fused into the recommendation algorithm. (2) A personalized recommendation algorithm based on context-aware and Zhang Liang decomposition is proposed and simulated on a large scale real data set of lens. The validity of the algorithm is verified. (3) based on the kernel density estimation technology and the context attribute information of users and items, the preference models of users and items are constructed, and a new similarity calculation method is proposed based on the preference model. Then the neighbor with high similarity measure is fused, and the user and project are recommended with certain recommendation methods.
【学位授予单位】：山东师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【参考文献】