融合用户信任关系和神经网络的推荐模型

A Recommendation Model Integrating User Trust Relationship and Neural Network

矩阵分解(Matrix Factorization, MF)是构建推荐系统的有效方法之一。然而,传统的矩阵分解模型在应用于大规模稀疏数据场景时依然存在推荐准确率较差、训练速度较慢等严重不足。本文从提高推荐准确率和减少时间花销这两方面考虑,提出一种基于神经网络(Neural Network, NN)的融合用户信任关系(Trust)的矩阵分解(NN-MF-TR)模型。一方面,通过用户间潜在的信任与被信任关系充分挖掘用户的潜在偏好,从而提高推荐准确率。另一方面,在训练模型时引入神经网络,减少因多次迭代所造成的巨大时间花销。最后,通过选取四个有代表性的对比模型并在四个真实世界数据集上进行了实验验证。实验结果表明,本文提出的NN-MF-TR模型能够在保证提高推荐准确率的同时大大减少时间花销。

Matrix factorization is one of the effective methods to construct recommendation system. However, the traditional matrix factorization model still faces the problems of both poor recommendation accuracy and low training speed when applied to the case with large-scale sparse data. In order to improve the recommendation accuracy while reducing the time cost, this paper proposes an NN-based matrix factorization model combining user trust relationships (NN-MF-TR). First, the la-tent preference of users is fully mined through the latent trust and trusted relationships between users, so as to improve the recommendation accuracy. Then, a neural network is introduced to train the constructed model to reduce the huge time cost caused by multiple iterations. Finally, four rep-resentative comparison models are selected and tested on four real-world data sets. The results show that the NN-MF-TR model proposed in this paper can greatly reduce the time cost while en-suring the recommendation accuracy.