Fake news detection: When complex problems demand complex solutions

Abstract

Fake news detection is one of the most challenging problems in today's information and communication systems. In this article we address the challenge of detecting the generation and spreading of misleading information in the specific scenario of rumours propagation and clickbait. We realise that the construction of the dataset used to study this kind of problems dramatically affects the performance of the model and, thus, its selection. Hence, we conduct experiments with two datasets of different complexity. In experiment A, by using a simple dataset with rumour propagation data from Twitter, we demonstrate that good performance scores can be obtained without relying on the high computational cost of hyper-parameters tuning. In experiment B, an approach with fewer parameters and computational layers is not suitable to study clickbait with a larger dataset featuring more complex dynamics. Information deluge clearly demands the automation of the procedures for information treatment and the adequate combination of natural language processing and machine learning techniques. As the underlying problem is very complex, there is a tendency to think that the solution must be a complex model, i.e. a model with a large number of parameters and hyper-parameters. Our results confirm this idea, and underline the importance of identifying the most appropriate model assumptions based on the type of dataset available in order to select and configure the machine learning algorithm.