融合多方面潜在特征和神经网络的推荐模型

Recommendation Model Based on Multi-aspect Latent Feature and Neural Network

目前,潜在因子模型被广泛用于推荐,现有的方法大多利用用户与项目之间的交互信息来学习潜在特征,然而,用户和项目的潜在特征可能是来自多个方面.同时,考虑到神经结构可以替代矩阵分解中内积的形式,模拟用户和项目之间的交互,本文提出了一种融合多方面潜在特征和神经网络的推荐模型.首先,将推荐系统建模为包含丰富语义的异构信息网络,然后利用元路径和异构skip-gram模型提取并学习不同方面的潜在特征;其次,结合注意力机制将这些特征向量加权融合;最后,将得到的用户和项目的全局向量表示送入到神经网络中以实现评分预测.本文模型在movielens数据集和豆瓣电影数据集上进行了实验,结果表明,该算法相比于传统仅基于单一方面的算法和不采用神经结构的算法具有更低的平均绝对误差和均方误差.

At present, the latent factor model is widely used in recommendation system.Most of the existing methods use the interaction information between users and items to learn the latent features.However, the latent features may come from different aspects, such as the directors and movie types in movie recommendation.At the same time, considering that neural structure can replace the inner product in matrix factorization, this paper proposes a recommendation model based on multi-aspect latent feature and neural network.First, the heterogeneous information network is constructed by the relationships between different entities, then, the meta-paths and heterogeneous skip-gram model are used to extract the latent features of different aspects.Finally, the global vectors are fed into the neural network to achieve the rating prediction.Experiments on two real datasets in this paper show that the algorithm has lower MAE and RMSE than the traditional algorithms which just based on single aspect and the algorithm without neural structure.