基于关系图邻接矩阵逼近的推荐系统

在基于关系图约束的推荐方法中,引入用户图(项目图)约束的目的是保持原始的高维用户表征空间(高维项目表征空间)与低维的隐性用户表征空间(隐性项目表征空间)之间用户关系(项目关系)的一致性.不同于传统的基于关系图Laplacian矩阵的一致性约束,本文提出一种基于关系图邻接矩阵逼近的推荐模型,从相似性空间一致性角度进行约束,在保持高维表征空间与低维隐性空间的一致性关系的同时,可以一定程度上避免局部过拟合问题.在EachMovie与MovieLens数据集上的实验结果验证了本文算法的有效性.

Graph adjacency matrix approximation based recommendation system

In graph based recommendation methods,the goal of the graph constraint is to preserve the consistency of user relationships (item relationships) between high dimensional user representation space (item representation space) and low dimensional latent user representation space.Instead of applying the traditional Laplacian matrix based consistency constraint,a graph adjacency matrix approximation based recommendation model is proposed.In essence,the matrix approximation plays a role of directly imposing a consistency constraint on the different similarity metric spaces.Thus,not only the consistency of the user relationships (the item relationships) from different representation spaces can be well preserved,but also,the local over-fitting problem can be avoided to some extent.Experimental results on EachMovie and MovieLens datasets show the effectiveness of the proposed method.