摘要：

根据微博水军活动的特点,提出一种基于隐空间代价敏感学习的半监督水军识别方法。从内容、行为、社交关系3个视角选取微博账户的22个特征,结合矩阵隐空间分解、代价敏感学习和社交关系正则技术,构造代价敏感的半监督最大间隔分类模型,并利用随机梯度 下降算法求解模型的线性复杂度。实验结果表明,该方法在准确率、召回率和F1指标上均优于SMFSR和L2-SVMs方法,并且具有接近线性的学习速度。

关键词: 水军识别, 矩阵分解, 代价敏感学习, 社交关系正则, 隐空间

Abstract:

According to the characteristics of microblog spammers,this paper proposes a semi-supervised spammer identification method based on cost-sensitive learning in latent space.Firstly,it selects twenty-two features of microblog account from perspectives of contents,activities and social relations.Then,it obtains latent account vectors using matrix factorization method and constructs a novel cost-sensitive semi-supervised classification model with the maximum margin theory in latent space.In addition,a social relation regularization from following behaviors is formulated on the model.Finally,it develops a linear-complexity algorithm for solving the model with the stochastic gradient descent method.Experimental results show that the proposed method outperforms existing methods significantly,such as SMFSR and L2-SVMS,in terms of the evaluation measures of accuracy,recall and F1 score.It also obtains nearly linear training speeds.

Key words: spammer identification, matrix factorization, cost-sensitive learning, social relation regularization, latent space

中图分类号: 

• TP393