融合发文时序特征的用户属性预测方法

User Attribute Prediction Method Fusing Post Temporal Features

现有的用户属性预测方法通常基于用户发文的语义特征,忽略了能够体现发文之间依赖关系的时序特征。针对此问题,提出一种融合发文时序特征的用户属性预测方法。该方法基于用户发文流,利用Word2Vec生成具有语义特征的发文向量,然后通过双向长短期记忆(Bidirectional Long Short Memory, Bi-LSTM)神经网络提取时序特征,最后输入全连接层和Softmax实现属性预测。实验结果表明,与未使用时序特征的属性预测方法相比,该方法具有较好的精确率和召回率。

Previous studies on attribute prediction are mainly based on semantic features of users’ postings, while the temporal features that present the dependent relationship between the postings are neglected. In view of this research gap, the attribute prediction method based on temporal features is proposed. On the basis of post flows, the vectors with semantic features are generated by Word2Vec, and the temporal features are extracted by Bi-directional Long Short Memory neural network. Thus, attribute prediction is achieved when the vectors fused with temporal features are sent to the fully connected layer and Softmax. Experimental results demonstrate that this method has higher performance on precision and recall than other attribute prediction methods.