Sexual predator detection in chats with chained classifiers

This paper describes a novel approach for sexual predator detection in chat conversations based on sequences of classifiers. The proposed approach divides documents into three parts, which, we hypothesize, correspond to the different stages that a predator employs when approaching a child. Local classifiers are trained for each part of the documents and their outputs are combined by a chain strategy: predictions of a local classifier are used as extra inputs for the next local classifier. Additionally, we propose a ring-based strategy, in which the chaining process is iterated several times, with the goal of further improving the performance of our method. We report experimental results on the corpus used in the first international competition on sexual predator identification (PAN’12). Experimental results show that the proposed method outperforms a standard (global) classification technique for the different settings we consider; besides the proposed method compares favorably with most methods evaluated in the PAN’12 competition.