基于用户反馈信息与用户信任度的商品推荐技术的研究

发布时间：2018-02-09 09:20

本文关键词： 个性化推荐协同过滤数据稀疏概率矩阵分解　出处：《内蒙古大学》2016年硕士论文　论文类型：学位论文

【摘要】：自互联网中的信息海量化以来,信息超负荷已经是所有网民即将或已经遇到的问题,而相关性是解决信息超负荷问题众多方法中的一种主流方法。个性化推荐在电子商务网站也逐渐凸显出其重要的作用。该技术在发现用户潜在需求的基础上,将不相关的信息或者项目过滤掉,最后主动向用户推荐满足其需求的项目,以达到缓解信息过载的目的。本文在以用户—项目评分矩阵为主要甚至唯一的数据信息基础的概率矩阵分解模型PMF中融入用户与好友的社会信任关系,将好友对用户的影响线性聚合起来,得到即考虑用户自身兴趣又考虑好友兴趣对其产生的影响的用户潜在特征向量,对由矩阵分解得到的用户潜在特征向量进行约束,同时会利用一个参数限制该约束的程度。并考虑项目的评论文本信息也能在一定程度上反映项目的属性,因此利用LDA模型处理评论文本信息得到项目的主题分布向量。再利用指数函数将矩阵分解得到的项目潜在特征向量映射到主题分布向量上来,对由矩阵分解得到的项目潜在特征向量产生约束,最后提出基于用户反馈信息与用户信任度的商品推荐模型(MRM)。实验中,采用梯度下降的方法对后验概率进行最大化,并同时优化用户潜在特征向量、项目潜在特征向量以及主题分布向量,得到最终的主题分布矩阵、用户潜在特征矩阵和项目潜在特征矩阵,最终获得用户的预测评分。本文在Yelp评论网站真实数据上进行实验,本混合模型在典型的衡量推荐准确度标准MAE和RMSE上分别提升了9.7%和8.7%,实验结果显示基于用户反馈信息与用户信任度的商品推荐模型的推荐结果优于概率矩阵推荐算法,并验证了该混合推荐模型在缓解数据稀疏性问题上的有效性。
[Abstract]:Since the quantification of information in the Internet, information overload has become a problem that all Internet users will or have encountered. The relevance is one of the main methods to solve the problem of information overload. Personalized recommendation is also playing an important role in e-commerce websites. This technology is based on discovering the potential needs of users. Filter out irrelevant information or items, and finally proactively recommend items that meet the user's needs, In order to alleviate the information overload, this paper integrates the social trust relationship between users and friends in the probability matrix decomposition model (PMF), which is based on the user-item scoring matrix, which is the main or even the only data information basis. The influence of friends on users is linearly aggregated to obtain user potential feature vectors that take into account both user's own interests and the effects of friends' interests, and constrain the user's potential Eigenvectors obtained by matrix decomposition. It also uses a parameter to limit the extent of the constraint and considers that the comment text information of the item can also reflect the properties of the item to a certain extent. Therefore, the topic distribution vector of the item is obtained by using the LDA model to process the comment text information, and then the item potential eigenvector obtained by the matrix decomposition is mapped to the topic distribution vector by using the exponential function. Finally, a commodity recommendation model based on user feedback information and user trust is proposed. In the experiment, the gradient descent method is used to maximize the posterior probability. At the same time, the user potential feature vector, the item potential feature vector and the topic distribution vector are optimized to obtain the final topic distribution matrix, user potential feature matrix and item potential feature matrix. Finally gets the user's forecast score. This article carries on the experiment on the Yelp comment website real data, The hybrid model is improved by 9.7% and 8.7 on the typical standard of recommendation accuracy, MAE and RMSE, respectively. The experimental results show that the recommended results of the model based on user feedback and user trust are superior to those of the probability matrix recommendation algorithm. The effectiveness of the hybrid recommendation model in alleviating the problem of data sparsity is verified.
【学位授予单位】：内蒙古大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.3

【相似文献】