Early Detection of Promotion Campaigns in Community Question Answering

As is the case with many social media websites, the Community Question Answering (CQA) portal has become a target for spammers to disseminate promotion information. Previous works mainly focus on identifying low-quality answers or detecting spam information in question-answer (QA) pairs. However, these works suffer from long delay since they all rely on the information of answers or answerers while questions have been displayed on the websites for some time and attracted certain user traffic. As a matter of fact, spammers on CQA platforms also act as questioners and involve promotion information in their questions. So if they can be detected as early as possible, the questions will not appear on the websites and affect legitimate users. In this paper, we design a framework for early detection of promotion campaigns in CQA based on only question information and questioner profile. First, we propose a novel sampling method for identifying the questions that contain promotion information, which compose the positive dataset. We also sample an unlabeled dataset of unsolved questions during a certain period of time. Then, we compare the characteristics of question information and user profiles between the two datasets, which are also used as features in the learning process. Finally, we apply and compare several PU (Positive and Unlabeled examples) learning algorithms to find positive examples in the unlabeled dataset. In our approach, no answer side information is needed, which means that it can detect spamming activities as soon as the question is posted. Experimental results based on about 0.7 million questions derived from a popular Chinese CQA portal indicate that our approach can detect questions related to promotion campaigns as effectively as but more efficiently than the state-of-the-art QA pair level detection methods.