基于深度神经网络和概率矩阵分解的混合推荐算法

Hybrid recommendation algorithm based on deep neural network and probabilistic matrix factorization

摘要:

针对个性化推荐中用户和项目描述信息未充分利用，用户评分矩阵数据集极端稀疏的情况，提出了基于深度神经网络和概率矩阵分解（PMF）的混合推荐算法。首先对用户和项目描述信息进行预处理，形成包含用户偏好特征的用户和项目特征集，再将各特征输入深度神经网络模型中进行训练。同时，利用概率矩阵分解模型根据用户评分矩阵通过最大后验估计优化得到潜在特征向量。然后通过对概率矩阵分解模型的用户和项目潜在特征向量以及深度神经网络模型的真实特征向量进行迭代更新，收敛得到融合用户和项目真实信息的潜在特征向量，最后利用该特征向量对用户进行个性化推荐。实验证明，本文算法较经典推荐算法以及前人算法在均方误差与平均绝对误差指标上均有改善，说明本文算法的有效性。

Abstract:

Aiming at the facts that user and project description information is not fully utilized in personalized recommendation and user score matrix data set is extremely sparse, a hybrid recommendation algorithm based on deep neural network and probabilistic matrix factorization (PMF) is proposed. Firstly, user and item description information is preprocessed to form user and item feature sets containing user preference, and then each feature is fed into the deep neural network model for training. At the same time, the probabilistic matrix decomposition model is used to optimize the potential eigenvectors based on the maximum posterior estimation of the user score matrix. Then the potential feature vectors of the probabilistic matrix model and the real feature vectors of the deep neural network model are iteratively updated to converge to the potential feature vectors that fuse the real information of the user and the item. Finally, this feature vector is used to make personalized recommendation to users. Experiments show that the proposed algorithm is better than the classical recommendation algorithm and previous algorithms in term of the mean square error and mean absolute error index, which shows the effectiveness of the proposed algorithm.