Evaluation of machine learning classifiers in ordinal multiclass fake news detection scenario

Abstract

[EN] This thesis intends to explore machine learning classifiers and techniques to address the problem of fake news detection. Prediction algorithms can generate different results in this problem due to variance in dataset labeling caused by ambiguity and subjectivity of semantic text. The LIAR Dataset was used in the experiments of this thesis. This dataset derived from PolitiFact fact-checking agency data which is composed of a 6-class ordinal labeling that places political statements in the range between completely false and completely true statements. The original experiment that created the dataset achieved 27.4% class accuracy using hybrid CNN and Bi-Directional LSTM networks. The main contribution of this work consists of evaluating simpler classifiers focusing on using different preprocessing and feature selection techniques when modeling metadata and text features. Furthermore, this work explores the ordinal characteristics of the class labels and uses simple binary classifiers in an ordinal ensemble method already established in the literature.