【作者】 张千帆；

【导师】 瞿彬彬；

【作者基本信息】 华中科技大学 ， 计算机软件与理论， 2020， 硕士

【摘要】 在当今信息大爆炸的时代,各式各样的信息令人目不暇接。这不可避免地导致信息过载,使得人们在检索自己感兴趣的信息时存在一定困难。因此个性化推荐系统的重要性不言而喻。现有的推荐算法大多数都基于协同过滤,而协同过滤算法的主要不足之处在于其难以应对数据稀疏性问题。此外,协同过滤算法仅利用用户与物品的交互信息进行建模,其所挖掘的隐式信息仍不够充分。大多数相关方法仅关注显式或隐式信息中的一种,导致推荐的准确度并不高。针对上述问题,提出了一种基于数据集成和深度协同交叉网络的推荐算法。首先利用用户对物品的评分以及物品的标签信息计算用户对于标签的偏好,据此预测用户对物品基于偏好的印象分。然后从已有数据中挖掘出用户的社交关系信息,具体做法是先计算用户相似度,随后为每个用户定义可信用户集,通过计算第三方评分者与该可信用户集的相似度,预测出用户对物品基于社交关系的印象分。最后将上述两种印象分进行加权平均,得出社交偏好综合分,实现数据集成。在网络结构方面,将现有的协同交叉网络与基于深度矩阵分解的神经网络进行融合。融合后的网络通过任务关系矩阵实现两个基础网络之间的知识迁移和共享,并以完全相同的用户集和互不相交的物品集作为两个基础网络桥接的工具。最终,在推荐结果不重复的前提下,两个基础网络在训练时能够相互协作。在推荐系统常用数据集MovieLens和Book-Crossing上进行实验,评估了算法在Top-K推荐中的命中率以及推荐列表排序质量。此外还评价了评分预测的效果,并验证所提出的方法在关于推荐的各项常用评价指标上均优于当下前沿的推荐算法。最后对方法中涉及的超参数进行了敏感性分析,并得出使方法能够发挥最佳性能的超参数的取值。更多还原

【Abstract】 In the age of information explosion,all kinds of information can make our head spin.This inevitably leads to information overload and makes it difficult for people to retrieve information they are interested in.Therefore,the importance of personalized recommendation system is self-evident.Most of the existing recommendation algorithms are based on collaborative filtering.The main disadvantages of collaborative filtering algorithms are their difficulties in solving the problem of data sparsity.Moreover,collaborative filtering algorithms only use the interaction between users and items for modeling,and the implicit information mined by them is still insufficient.Most of the related methods only focus on one of the explicit and implicit information,which results in the poor recommendation accuracy.To solve the problems above,a recommendation algorithm based on data integration and deep collaborative cross network is proposed.First,the user’s preferences for tags are calculated by utilizing the user’s ratings and tags attached to items,and predicts the preference-based impression score of users for items.Then the user’s social relationship information is mined from the existing data.The concrete approach is to calculate the similarity among different users,and then the trusted user set is defined for every user.The social-relationship-based impression score of users for items are predicted by calculating the similarity between the third-party scorer and the trusted user set.Finally,the two kinds of impression score are weighted and averaged to obtain the comprehensive rating of social and preference to implement data integration.In terms of network structure,the existing collaborative cross network is fused with the neural network based on deep matrix factorization.The fused network implements the knowledge transfer and sharing between the two base networks via the task relation matrix,and bridges the source network and the target network through the identical user set and the mutually disjoint item set.Finally,under the premise of non-repetitive recommended results,the two basic networks can cooperate with each other in their training.The experiments are conducted on MovieLens and Book-Crossing data sets which are widely used for recommendation systems.It evaluates the hit ratio in top-K recommendation and the sorting quality of recommendation list.In addition,the performance of rating prediction is evaluated,and it is proved that the metrics on the tasks of score prediction and top-K recommendation of the proposed method are all better than the current state-of-the-art recommendation algorithm.Finally,the sensitivity of the hyper-parameters involved in the method is analyzed by experiments,and the values of the hyper-parameters that make the method perform best are obtained.更多还原