ACTSSD: social spammer detection based on active learning and co-training

• Published in: The Journal of supercomputing 2022-02, Vol.78 (2), p.2744-2771
• Main Authors: Chen, Ailin, Yang, Pin, Cheng, Pengsen
• Format: Article
• Language: English
• Subjects: Active learning; Algorithms; Compilers; Computer Science; Datasets; Interpreters; Labels; Machine learning; Processor Architectures; Programming Languages; Social networks
• ISSN: 0920-8542; 1573-0484
• DOI: 10.1007/s11227-021-03966-3

The rumors, advertisements and malicious links are spread in social networks by social spammers, which affect users’ normal access to social networks and cause security problems. Most methods aim to detect social spammers by various features, such as content features, behavior features and relationship graph features, which rely on a large-scale labeled data. However, labeled data are lacking for training in real world, and manual annotating is time-consuming and labor-intensive. To solve this problem, we propose a novel method which combines active learning algorithm with co-training algorithm to make full use of unlabeled data. In co-training, user features are divided into two views without overlap. Classifiers are trained iteratively with labeled instances and the most confident unlabeled instances with pseudo-labels. In active learning, the most representative and uncertain instances are selected and annotated with real labels to extend labeled dataset. Experimental results on the Twitter and Apontador datasets show that our method can effectively detect social spammers in the case of limited labeled data.