基于改进的宽深度模型的推荐方法研究

THE RECOMMENDATION METHOD BASED ON IMPROVED WIDE AND DEEP MODEL

现代社交网络的个性化博文推荐中,博文特征选取质量的高低直接影响了推荐的质量和效率。深度模型可以较高质量地提取出文本中句法和语义的特征。然而短文本特征稀疏且未考虑上下文语境的问题,普遍存在于文本推荐任务中。针对以上问题,在现有宽深度模型的基础上,利用门限循环单元对其多层普通神经网络进行改进,提出宽深度门循环联合(Wide&Deep-GRU)模型,进一步探索浅层部分和深度部分的联合训练。使用从新浪微博获取的真实数据集分别与单一逻辑回归模型、单一深度神经网络模型和宽深度模型进行对比。实验表明,该方法整体上推荐质量较高,同时推荐效率较之前模型也有显著提高。

In the personalized blogs recommendation of modern social networks,the quality of blog feature selection directly affects the quality and efficiency of recommendation.Deep model can extract syntactic and semantic features from text in a higher quality.However,the short text features are sparse and do not consider the context,which are ubiquitous in the text recommendation tasks.Aimed at above problem,a wide and deep-GRU model,based on the existing wide and deep model,was proposed by using Gated Recurrent Unit to improve its ordinary multilayer neural networks so as to further explore the joint training of the shallow part and depth part.In this paper,the real datasets obtained from Sina Weibo were respectively compared with single logistic regression model,single depth neural network model and the wide and deep model.Experiments show that this method generally achieves a good recommendation effect and the recommended efficiency is significantly higher than the previous model.