基于排序学习和卷积神经网络的推荐算法研究

发布时间：2018-06-03 14:42

本文选题：推荐系统 + 社交网络　；参考：《大连理工大学》2016年硕士论文

【摘要】：随着互联网技术特别是以淘宝和亚马逊等为代表的电子商务的飞速发展,互联网中的数据呈现爆炸性增长,信息过载问题显得越来越严重。帮助我们从海量数据中筛选出有意义数据的信息过滤技术显得越来越重要。在此背景下,推荐系统诞生了,并且迅速发展成为当前互联网应用中的重要组成部分。推荐系统根据用户行为记录从大规模数据中找到用户感兴趣商品,它对于提高用户的满意度和零售商的销售额具有重要的意义。用户在互联网中的行为主要分为两类,分别是隐性反馈行为和显性反馈行为。其中在隐性反馈行为中用户没有显式地表达对特定商品的偏好,主要包括用户的点击、浏览、收藏等行为；而在显性反馈行为中用户则显式地表达了对特定商品的偏好信息,这些行为中较为常见的主要有评分行为。针对不同类型的用户反馈行为数据有不同的推荐方法,本文对两种不同的用户反馈行为进行了细致地分析和挖掘,并且分别有针对性地提出了两种方法以提高推荐系统的性能。针对显性反馈行为的评分行为,本文选取Top-K推荐作为研究目标。引入信息检索领域排序学习的方法并且融合用户的社交信息和商品标签信息,本文扩展了一种基于列表排序学习的矩阵分解方法,一方面充分考虑用户之间关注关系。首先通过用户之间的关注关系计算用户之间的信任度,接着通过用户之间的信任度在原始模型的损失函数中添加用户社交约束项,使相互信任的用户偏好向量尽可能接近。另一方面,计算商品所拥有标签的权重并以此计算商品之间的标签相似度,再将商品的标签约束项添加至损失函数中。在真实Epinions和百度电影数据集中的实验结果表明,我们提出的方法的NDCG值和原始模型相比具有一定的提高,有效地提高了推荐准确率。针对隐性反馈行为,本文选取电子商务领域的下一个购物篮推荐作为研究目标。本文首先将用户行为按照一定的时间窗口进行划分,对于每个窗口从多个不同的维度抽取用户对商品的时序偏好特征；接着运用深度学习领域的卷积神经网络模型,模型中的卷积层组合不同长度的特征图来训练分类器。在阿里巴巴移动推荐算法竞赛公布的真实数据集中的实验结果表明,和传统的线性模型和树模型等分类器相比,我们提出的卷积神经网络框架具有较强的特征萃取能力和泛化能力,提高了推荐系统的用户满意度。
[Abstract]:With the rapid development of Internet technology, especially the electronic commerce, such as Taobao and Amazon, the data in the Internet is growing explosive. The problem of information overload is becoming more and more serious. Information filtering techniques that help us filter meaningful data from mass data are becoming more and more important. In this context, the recommendation system is in the background. The system is born, and has rapidly developed into an important part of the current Internet applications. The recommended system is based on user behavior records to find users interested in goods from large-scale data. It is important to improve the satisfaction of users and the sales of retailers. The behavior of users in the Internet is divided into two categories. There is no implicit feedback behavior and explicit feedback behavior. In the implicit feedback behavior, users do not explicitly express preference for specific goods, including users' click, browse, collection and other behaviors, while in explicit feedback behavior, users express preference information about specific products, which are more common in these behaviors. There are different methods of recommendation for different types of user feedback behavior data. In this paper, two different user feedback behaviors are carefully analyzed and excavated, and two methods are proposed to improve the performance of the recommended system respectively. In this paper, the paper selects Top for the behavior of dominant feedback behavior. -K recommends as a research goal. Introducing the method of sorting learning in the field of information retrieval and integrating the user's social and commodity label information, this paper extends a matrix decomposition method based on list sorting learning. On the one hand, it takes full consideration of the concerns between users. And then the user's social constraints are added to the loss function of the original model through the trust degree between the users, so that the mutual trust user preference vector is as close as possible. On the other hand, the weight of the label is calculated and the label similarity between the goods is calculated, and the label constraint item of the commodity is added to the loss. In the loss function, the experimental results in the real Epinions and Baidu movie datasets show that the NDCG value of the proposed method is improved to a certain extent compared with the original model, which effectively improves the accuracy of the recommendation. In this paper, the next shopping basket in the field of electronic commerce is selected as the research goal. First, the user behavior is divided according to a certain time window, and each window is extracted from a number of different dimensions of the user's timing preference. Then, the convolution neural network model in the depth learning field is used to train the classifier with different length of feature graph to train the classifier. In the Alibaba movement, the classifier is moved and pushed. The experimental results of the true data set published in the recommendation algorithm contest show that, compared with the traditional linear and tree model classes, the convolution neural network framework proposed by us has strong feature extraction ability and generalization ability, and improves the user satisfaction of the recommendation system.
【学位授予单位】：大连理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.3;TP183

【参考文献】