基于深度学习和社交关系正则化的混合协同过滤推荐算法

发布时间：2018-06-06 14:17

本文选题：推荐系统 + 深度学习　；参考：《广东工业大学》2017年硕士论文

【摘要】：随着互联网的快速发展,网络信息呈现爆炸式的增长,其结构也变得越加复杂。海量信息的呈现,使得用户很难从中发现自己感兴趣的内容,而推荐系统可以帮助用户发掘更深次的需求,给用户带来个性化的体验。此外,推荐系统可以帮助用户更容易地找到他们需要的产品,也可以通过改进用户体验帮助企业提升用户忠诚度从而把更多的潜在用户转换为产品购买者。同时,推荐系统也具有研究价值,涉及计算数学,认知科学,信息科学等学科。在推荐系统中,协同过滤是目前应用最广泛的一种个性化推荐技术。传统的协同过滤方法仅仅使用用户对物品的评分矩阵进行推荐。在实际情况中,通常用户的评分矩阵非常稀疏,从而导致推荐效果不佳。在这种情况下,一些模型尝试使用物品内容信息来缓解数据稀疏和冷启动问题。然而,当这些内容信息也非常稀疏时,很难从这些模型中学习到准确的特征表示进行推荐。为了应对这些问题,本文提出了基于深度学习和社交关系正则化的混合协同过滤推荐模型(CDL-SR)。其中,社交正则化在推荐系统中表示一种社交约束。该模型利用深度学习强大的特征表达能力,将通过深度学习算法自动学习到的物品特征表达向量同矩阵分解后的评分矩阵有效融合,并通过在目标函数中加入社交正则项,让存在社交关系的物品(如存在引用关系的论文)的隐式特征向量具有较高相似度以进一步提高推荐效果。C D L-S R模型不仅可以提供新用户的个性化推荐信息(冷启动),还有利于解决用户评分矩阵及物品的文本、属性、码流等信息的数据稀疏问题。本文以Cite Ulike论文集为样本进行实验研究表明,本文采用的深度学习与社交网络信息相结合的方法能够提供更好的推荐性能。特别是在稀疏数据集下,该方法相比于目前流行的协同主题回归模型(CTR),召回率提升了66.7%。此外,该推荐系统在推荐结果中可以给出较为准确和令人信服的推荐理由,进一步提高了用户对系统的满意度。
[Abstract]:With the rapid development of the Internet, the network information explosive growth, its structure has become more and more complex. The presentation of mass information makes it difficult for users to find the content they are interested in, and recommendation system can help users to explore deeper needs and bring personalized experience to users. In addition, recommendation systems can help users find the products they need more easily, and can also help businesses improve their customer loyalty by improving their user experience, thus transforming more potential users into product buyers. At the same time, recommendation system also has research value, involving computational mathematics, cognitive science, information science and other disciplines. Collaborative filtering is the most widely used personalized recommendation technology in recommendation systems. The traditional collaborative filtering method only uses the user's scoring matrix to recommend items. In practice, the user's rating matrix is usually very sparse, resulting in poor recommendation results. In this case, some models attempt to use item content information to alleviate data sparsity and cold startup problems. However, when the content information is sparse, it is difficult to learn the exact feature representation from these models to recommend. In order to solve these problems, a hybrid collaborative filtering recommendation model based on deep learning and social relationship regularization is proposed in this paper. Social regularization represents a social constraint in a recommendation system. Using the strong feature expression ability of depth learning, the model effectively integrates the feature expression vector of objects automatically learned by the depth learning algorithm with the score matrix after matrix decomposition, and adds social regular items to the objective function. The implicit feature vectors of objects with social relationships (such as papers with reference relationships) have high similarity so as to further improve the recommendation effect. The CD-L-S R model can not only provide personalized recommendation information for new users ( Cold boot can also help to solve the user rating matrix and the text of the item. Attribute, bitstream and other information sparse problem. In this paper, the experimental results show that the combination of in-depth learning and social network information can provide better recommendation performance. Especially in sparse data sets, the recall rate of this method is 66.7% higher than that of CTRN, a popular cotopic regression model. In addition, the recommendation system can give more accurate and convincing recommendation reasons in the recommendation results, which can further improve the satisfaction of users to the system.
【学位授予单位】：广东工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【参考文献】