基于Co-training的用户属性预测研究

发布时间：2018-06-10 06:59

本文选题：用户属性 + Co-training　；参考：《工程科学与技术》2017年S2期

【摘要】：针对当前基于第三方应用数据进行用户属性预测算法研究,其较少考虑应用前台实际使用时长问题,由此,本文在应用的使用频率及使用时长的基础上,构造了应用前台均使用时长特征,该特征能进一步刻画用户对应用的兴趣度;同时,为充分利用大量未标注数据,从多角度特征对用户属性进行预测,由此本文采用了Co-training框架,该框架包含两个均由栈式自编码器与神经网络相结合的网络结构。实验过程中,对于栈式自编码算法,先利用未标注的数据对网络进行参数初始化,使得网络参数处于一个较优的位置,再利用有标注的数据,采用基于准确率的梯度下降算法,对网络参数进行更新,最终达到收敛。实验结果表明,本文算法在准确率、召回率、F1值上均有所提高。
[Abstract]:In view of the current research on user attribute prediction algorithm based on third-party application data, the problem of actual usage time of application foreground is less considered. Therefore, based on the frequency and duration of application, In order to make full use of a large amount of unannotated data and to predict user attributes from multiple angles, the Co-training framework is used in this paper. The framework consists of two networks which are composed of stack self-encoder and neural network. In the process of experiment, for the stack self-coding algorithm, the network parameters are initialized with unlabeled data at first, and the network parameters are placed in a better position. Then, using labeled data, the gradient descent algorithm based on accuracy is adopted. The network parameters are updated and finally converged. The experimental results show that the accuracy and recall rate of the algorithm are improved.
【作者单位】：四川大学计算机学院;
【基金】：国家自然科学基金资助项目(61332066;81373239)
【分类号】：TP301.6

【相似文献】