Unsupervised Learning For Spammer Group Detection Based On Network Representation

Nowadays,we are used to get all kinds of information from the Internet with the development of the Internet.However,the information on the Internet may not be true.The spammers make huge damage to economic and social environment and have a terrible influence on the network.Spammer detection is an necessary job for the development of Internet platform healthy.However,because spammers are professional and they always imitate the behavior and language of normal users,which makes the detection effect of the traditional spammer identification method getting worse.It is hard to distinguish spammers from all users by reading the reviews.Moreover,the spammers will constantly adjust their behavior according to the platform governance rules.As a result,The spammer detection method needs to change constantly.Supervised detection models always need many people to get labels and these labels may be biased.The existing unsupervised models for spammer detection have low accuracy.To solve the problem,we consider that the spammers are becoming organized.So we focus on the most essential and unchangeable characteristics of spammer groups to propose a novel unsupervised spammer group detection method.To influence more users,spammer groups need to create public opinion firstly.So a large number of spammers will review the same postings with similar standpoint in a short time.Which is different from normal users obviously.Our aim is detecting these organization spammer groups which have terrible influence on the internet and ignoring individual spammers.According to this,our method based on the fact that the spammer groups in online media or shopping websites connected closely and their reviews are similar while normal users have weak connections and low similarities.Firstly,we established a user network based on vari-ous features such as comment content similarity and user behavior similarity.Then we proposed a new community detection method based on network representation learning to detect closely-connected communities.Finally,we used some indicators to identify spammer groups.By experimenting on four real-world datasets,we prove the proposed model proposed in this paper is superior to both traditional and state-of-the-art unsupervised spammer de-tection methods.