一种优化的带偏置概率矩阵分解算法

Optimization Algorithm of the Probabilistic Matrix Factorization with Bias

针对个性化推荐过程中高维稀疏性引起的数据震荡和推荐精度不高的问题,提出一种通过交替最小二乘算法(Alternating Least Squares,ALS)来优化的带偏置概率矩阵分解的推荐方法.首先将用户项目的偏置信息融入到改进的概率矩阵分解算法中.其次为了提升训练速度和推荐精度,将训练得到的用户项目潜在因子向量作为ALS的初始值,进而得到用户项目潜在因子矩阵.最后利用分解后的两个低维矩阵对原矩阵中的未知评分进行预测.在Movielens100k数据集上的实验结果表明,本文提出的推荐算法在相对于传统的带偏置概率矩阵分解来说最高提高3.41%,结果稳定且准确率高.

Concerning the problem of the situation that the performance fluctuate greatly and low recommendation accuracy during personalized recommendation,this paper proposes an improved probabilistic matrix factorization by alternating least squares algorithm. Firstly the bias information is merged into the traditional probabilistic matrix factorization. Secondly, we take the trained potential factor vector from the traditional probabilistic matrix factorization as the initial value of alternating least squares to get the new potential factor vector for the sake of improving the training speed and the accuracy of the recommendation. Finally we use the new potential factor matrix of users and items to recommend. The results on the Movielens100k dataset show that the proposed algorithm can effectively and stably improve the recommendation accuracy by 3.41.