基于SDAE特征表示的协同主题回归推荐模型

A collaborative topic regression recommendation model based on SDAE feature representation

为解决推荐系统中的冷启动问题,在协同主题回归CTR模型的基础上引入堆叠去噪自编码器SDAE深度学习网络,用于学习用户辅助信息的隐表示,建立SDAE-CTR模型。模型应用2层SDAE网络,以用户信息为网络输入量,将编码过程获得的用户辅助信息的隐表示和解码过程获得的输入近似表示为网络的双输出量,最小化用户辅助信息和近似表示的差值来确定最优隐表示。模型融合用户-项目评分矩阵(冷启动条件无评分)、项目内容信息和用户辅助信息实现用户对未评分项目的评分预测,并在LastFM、Book Crossing和MovieLens数据集上从推荐准确度、新颖性和用户冷启动条件下的推荐效果等3方面对SDAE-CTR模型和CTR模型进行比较。结果表明,SDAE-CTR模型在冷启动或非冷启动的条件下,推荐效果都要优于CTR模型的,虽然新颖性较CTR模型稍微逊色一些,但理论上在合理的范围内,总体上SDAE-CTR模型表现较优。

In order to address the cold start problem in the recommendation system,based on the collaborative theme regression (CTR) model,we introduce the stacked denoising autoencoder (SDAE) deep learning network to deeply learn the implicit representation of user auxiliary information,and establish a SDAE-CTR model.The model utilizes a two-layered SDAE network,takes the implicit representation obtained in the encoding process and the approximate representation obtained in the decoding process as the two network outputs,and determines the optimal implicit representation by minimizing the difference between the user auxiliary information and the approximate representation.The model utilizes a user-item scoring matrix (cold start condition without scoring),item content information and user auxiliary information to predict user ratings for unrated items.We compared the SDFM-CTR to the CTR model in recommendation accuracy,novelty,and recommendation effect under the user cold-start condition on the LastFM,Book Crossing,and Movie Lens datasets.The results show that the SDAE-CTR is better than the latter under the conditions of cold start or non-cold start,the degree of novelty is theoretically within a reasonable range,and its overall performance is better.