用于个性化推荐的条件卷积隐因子模型

在推荐系统中,传统的矩阵分解无法提取用户和物品特征,而神经协同过滤(NCF)在分解模型中增加多层感知器,但不能有效利用用户和物品ID之外的辅助信息.为此,提出一种新的条件卷积方法.通过将物品特征作为输入,将用户特征作为卷积核,达到权值不共享的目的,使得条件卷积具有更强的特征提取和组合能力以及不增加参数量的特性.在此基础上,条件卷积能够融入多种辅助信息进行个性化推荐.实验结果表明,与NCF模型相比,该方法在隐性反馈数据中推荐命中率提升3.11%,在显性反馈数据中评分预测误差降低2.47%.

Conditional Convolution Implicit Factor Model for Personalized Recommendation

In the recommendation system,the traditional Matrix Factorization(MF)cannot extract the features of users and items.The Neural Collaborative Filtering(NCF)adds a multi-layer perceptron to the model for this reason,but still cannot effectively make use of the auxiliary information other than the user and item ID.Therefore,this paper proposes a new conditional convolution method.In this method,the item features are used as input and the user features as convolutional kernels,so as to achieve the purpose of unshared weights,thus making the conditional convolution has stronger feature extraction and combination abilities while maintaining the features of parameters.On this basis,the conditional convolution can incorporate a variety of auxiliary information for personalized recommendation.Experimental results show that compared with the NCF model,the proposed method can increase the recommendation hit rate by 3.11%in the implicit feedback data and reduce the scoring prediction error by 2.47%in the explicit feedback data.