Abstract:

Most traditional recommendation algorithms only consider user-commodity rating information for recommendation, so that the accuracy of these recommendation models are not high, which ignore the information of user attributes and commodity attributes. Therefore, a hybrid recommendation model based on deep learning（Deep and Cross Factorization Machine, DCFM） is proposed by using factorization machine to mine the relationship between users and commodities in the case of sparse data, cross-network to mine linear combination relation between the first input layer and every high order features of attributes and deep neural network to mine high-order non-linear association. The three parts are connected in parallel, sharing the input layer, and the weighted results are combined by linear stitching as the overall output of the model. Through the simulation experiment on MovieLens movie data set, and compared with Factorization Machine（FM）, Deep Factorization Machine（DeepFM）, Deep Cross Network（DCN） model, the results show that the model has been improved in accuracy, F1-Score and AUC values.

Key words: intelligent recommendation, deep learning, factor machine, cross network

摘要：

传统推荐算法大多都仅考虑用户-商品评级信息来进行推荐，这种忽略了用户属性和商品属性信息的推荐模型准确率不高。因子分解机可在数据稀疏情况下挖掘用户与商品的关联关系，交叉网络可挖掘属性特征与其高阶特征的线性组合关系，以及深度神经网络有效识别高阶非线性关联关系，基于三种模型的优势，提出了一种基于深度学习的混合推荐模型（Deep and Cross Factorization Machine，DCFM）。三部分并联组合，共享输入层，各部分结果线性组合后作为模型整体输出。通过在MovieLens电影数据集上仿真实验，并与因子分解机（FM）、深度因子分解机（DeepFM）、深度交叉网络（DCN）模型做比较，结果证明该模型在准确率、F1-Score和AUC值上均得到了提高和改善。

关键词: 智能推荐, 深度学习, 因子分解机, 交叉网络