Research shows that microblogging websites such as Tumblr are being misused as a platform to disseminate hate and extremism. We formulate the problem of locating such extremist communities as a graph search problem. We propose a topical crawler based approach performing several tasks: searching for a blogger, computing its similarity against exemplary documents, filtering hate promoting bloggers, navigating through links to other bloggers and managing a queue of such bloggers for social network analysis. We conduct experiments on real world dataset and examine the e↵ectiveness of 'like' and 'reblog' features as links between bloggers. Experimental results demonstrates that the proposed solution approach is e↵ective with an F-score of 0 .80.
1. PROBLEM DEFINITION & SOLUTION
Tumblr is a popular and widely-used micro-blogging
website. Previous research shows that such websites are used
as a platform for disseminating hate and extremism (due
to low barrier to publication and anonymity) [
Copyright c 2015 held by author(s)/owner(s); copying permitted only for private and academic purposes.
Published as part of the #Microposts2015 Workshop proceedings, available online as CEUR Vol-1395 (http://ceur-ws.org/Vol-1395)
In a graph traversal, a topical crawler returns relevant nodes to a specific topic. To define the relevance of a node, it learns the characteristics and features of given topic and computes the extent of similarity against a bunch of exemplary documents. To collect training examples, we perform an iterative search on Tumblr using keyword based flagging, where keyword is a search tag; for example, jihad, anti-Islam and hate. We perform a case study on Jihad and by manual search on Tumblr posts we collect several relevant tags that are commonly used by extremist bloggers. We use these tags to initiate our process and collect all textual posts (avoiding picture, audio, video and URLs), tags (associated with resultant posts) and linked bloggers (post reblogged by and liked by) with no redundancy. We perform a manual inspection on resultant posts and posts made by linked bloggers to filter relevant (hate promoting) and unknown results. We further extract more posts and linked bloggers from related tags and run this framework recursively to collect our exemplary documents (400 hate promoting posts). These training examples contain the body and caption of only positive class (hate and extremism promoting content) posts which is used to train the model.
Figure 1 illustrates the design and architecture of topical crawler to locate extremist communities. As shown in Figure 1, our proposed solution framework is an iterative multi-step process primarily consisting of five phases: features (posts) extraction, data pre-processing, classification, frontier extraction and graph traversal. In phase 1, we initiate our process using a positive class (hate promoting) blogger Ui called as ’seed’. We use Tumblr API 1 to fetch the URLs of n number of textual posts and by using Jsoup Java library 2 we extract the content and caption of these posts (used as contextual metadata). These posts can be either re-blogged from other users or originally posted by the user Ui. These posts consist of multiple langauges. Therefore, in phase 2, we perform data pre-processing and filter English and non-English posts using language detection library3. We perform data pre-processing on these posts and remove English stopwords. In phase 3, we build a statistical model from the exemplary documents collected separately by semi-automatic process. To compute the relevance of each blogger, we use character level n-gram language modeling approach. We find the extent of similarity between metadata and exemplary documents using LingPipe API 4 - applying joint probabilitybased classification of character sequences . We implement a one class classifier and filter extremism promoting bloggers from unknown bloggers. In phase 4, we extract the notes associated with the posts (collected in phase 1) of relevant bloggers. These notes contain the list of bloggers who liked and re-blogged a particular post. The number of notes represent the popularity of a post and indicate the similar interest between original poster and other bloggers in the list who may or may not be the direct followers of each other. We use notes to extract frontier nodes of a blogger because of two reasons: 1) due to the privacy policies Tumblr API does not allow developers to extract followers and following blogs of Tumblr users. 2) Tumblr facilitates bloggers to track any number of tags so that whenever there is a new post published publicly on Tumblr containing any of these tags, it automatically appears in a menu on user’s dashboard. They can spread that post among their followers by re-blogging it. Tracked tags allow bloggers to form a virtual community without following each other. For each frontier extracted in phase 4, we compute the relevance score against exemplary documents and discard unknown bloggers. In phase 5, we manage a queue of relevant bloggers and perform directed graph traversal using random walk algorithm. To expand our graph we select the next blogger in uniform distribution and extract it’s frontiers. We execute our focused crawler for each frontier without revisiting a blogger. This traversal results in a connected graph, where nodes represents a blogger (hate promoting) and edges represent the links (re-blog and like) between two bloggers. We perform social network analysis on the resultant graph and locate extreme right communities of hate promoting bloggers. 2. RESULTS & CONCLUSION We execute our topical crawler for a given seed blogger and traverse through Tumblr network using random walk algorithm. For every new blogger, we compute its relevance and classify it as hate promoting or unknown using one class classifier. To examine the e↵ectiveness of our classifier, we compute its accuracy using standard information retrieval techniques. In one execution of our topical crawler, we were able to collect 600 bloggers. We hired 30 graduate students as volunteers from die↵rent department to label these bloggers as hate promoting or unknown according to their published posts and given guidelines for annotation. To avoid the biasness and to collect correct annotated results we perform a horizontal and vertical partition on nodes and arrange these 600 bloggers into a 2D matrix where rows are the numbers of annotators grouped in 10 sets, 3 members each. Columns of the matrix are the number of bloggers assigned to each member for annotation i.e. 60. We use majority voting approach for final annotation, the class of a blogger is the one which is voted by at least two annotators. Based upon the validation results we evaluate the accuracy of our model. Table 1(a) shows the confusion matrix for one class classification. Table 1(a) reveals that our model predicts 382 (290+92) bloggers as hate promoting and 218 (173+45) bloggers as unknown. Table 1(a) shows that there is a missclassification of 13% and 34% in predicting hate promoting and unknown bloggers. Table 1(b) shows the accuracy results of our classifier. Results shows that the precision, recall and f-score are reasonably high and we are able to predict hate promoting bloggers with an accuracy of 77%. Our experimental analysis reveals that re-blogging is a good indicator of connection between two bloggers. We locate users who are central and influential among all and play major role in the discovered communities. We perform independent social network analysis on like and re-blog links among bloggers and conclude that re-blogging is a discriminatory feature to identify the communities of extremist bloggers sharing a common agenda.