一种可解释性泛化矩阵分解推荐算法

An explainable generalized matrix factorization recommendation algorithm

可解释性矩阵分解解决了概率矩阵分解缺乏可解释性的问题,然而概率矩阵分解假设评分数据是服从正态分布的,这与实际场景有一定的偏差.针对该问题,提出一种可解释性泛化矩阵分解推荐算法.首先采用一种新型的变换函数使原始评分近似服从正态分布,然后通过可解释性矩阵分解获得预测评分,最后利用对应的逆变换函数将预测评分映射回原始评分区间.在三个数据集上进行实验,结果表明,与多个主流矩阵分解算法相比,提出的算法在多个评价指标上占优.

Explainable matrix factorization solves the problem of the lack of explainability of probabilistic matrix factorization.However,this method assumes that the ratings obey the normal distribution,which is different from the real scenario.In this paper,we propose an explainable generalized matrix factorization recommendation algorithm.First,we use a novel transformation function to make the original ratings obey the normal distribution.Then,we predict ratings by explainable matrix factorization.Finally,we use the corresponding inverse transformation function to map the predicted ratings back to the original rating interval.The superiority of our method is substantiated by experiments on three datasets,as compared with state-of-the-art matrix factorization methods.