Predicting Tag Spam Examining Cooccurrences, Network Structures and URL Components

The task of the ECML/PKDD Discovery Challenge 2008 is to identify spammers in a social bookmarking system. We classify users using three different types of features, based on cooccurences, network properties and url parts. Cooccurrence features are based on the assumption that users associated with similar documents and tags as spammers are likely to be spammers themselves. Network-based features work on a collective scale, assuming common behavioural patterns which can be identified in the graph structures created by tagging activities. Finally, a text classification on the URLs’ components identifies frequent terms in spam URLs. With these features, we train an SVM for classification. Our submission run, combining all three classes of features, performed worse than expected from previous tests. With the wisdom of hindsight, we find an optimal choice of features is to leave out network features entirely but to strengthen URL classification. This is, however, a side effect of wrong assumptions about the test set; network features, used alone, still yield positive results. As network features do not depend on the presence of labeled users, they should be further explored to identify structural properties of tag spam even when no ground truth exists.