Automated analysis of pedophilic conversations using Machine Learning

Abstract

This research project is based on implementing a software solution to detect text messages with pedophilic intentions through mobile messaging applications. A corpus of this type of conversations is analyzed to select the most relevant features using natural language processing techniques and achieve a better performance in the predictive model with supervised Machine Learning classification algorithms. The Support Vector Machine algorithm was adopted as the text classification model, with the tests performed on this model very promising results were obtained with an accuracy of 80% in the classification of these messages, some of these had a lot of noise and grammatical errors so that the learning capacity of the model was affected, Therefore, after a processing stage of the data set and the necessary adjustments, it was possible to optimize the model a little more and reach 84% in terms of precision and accuracy, in addition to the F1 score that indicates a performance of 86% for the classification model built. Finally, the model is implemented in a Bot that is added to a Telegram group by connecting to its API for test conversation analysis to observe how well it performed the automatic classification of messages sent in the group.

Keywords: Online grooming, Machine learning, Automatic classification, Mobile messaging.