Proposta e avaliação de mecanismos de combate à poluição em sistemas de compartilhamento de vídeos

Abstract

Most online video sharing systems (OVSSs), such as YouTube and Yahoo! Video, have several mechanisms for supporting interactions among users. One such mechanism is the video-response feature in YouTube, which allows a user to post a video in response to another video. While increasingly popular, the video-response feature opens the opportunity for non-cooperative users to introduce 'content pollution' into the system, thus causing loss of service effectiveness and credibility as well as waste of system resources. For instance, non-cooperative users, to whom we refer as spammers, may post unrelated videos in response to another video (the responded video), typically a very popular one, aiming at gaining visibility towards their own videos. In addition, users referred to as content promoters post several unrelated videos in response to a single responded one with the intent of increasing the visibility of the latter. Previous work on detecting spammers and content promoters on YouTube has relied mostly on supervised classification methods. The drawback of applying supervised solutions to this specific problem is that, besides extremely costly (in some cases thousands of videos have to be watched and labeled), the learning process has to be continuously performed to cope with changes in the strategies adopted by non-cooperative users. In this work, we explore the use of multi-view semi-supervised strategies, which allows us to reduce significantly the amount of training, to detect non-cooperative users on YouTube, while keeping high levels of effectiveness. Our proposed method explores the fact that, in this problem, there is a natural partition of the feature space in sub-groups or views'', each being able to classify a given user when enough training data is available. Moreover, we propose to deal with the problem of view combination as a rank aggregation problem, where rankings based on confidence in the classification are combined to decide whether an unlabeled example should be included in the training set. Our results demonstrate that we are able to reduce the amount of training in about 80% without significant losses in classification effectiveness. Finally, we develop an analytical model to estimate the costs associated with the utilization of different methods to identify non-cooperative users in OVSSs. We here apply this model in different scenarios in order to compare our best proposed method (a hybrid method) with a supervised method which uses all the training data available (our baseline). The results of this analysis showed that our method has a lower cost when compared to the baseline for most of the analyzed scenarios.