Publication: FAKE NEWS DETECTION MODEL FOR THE EARLY STAGES OF THE SPREAD
Date
2021-04
Authors
ESPINOZA VILLARROEL, IGNACIO JAVIER
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
El masivo uso de redes sociales ha permitido un aumento explosivo de noticias falsas circulantes en la red. La principal razón es que este tipo de contenido puede ser creado y publicado rápidamente a un costo nulo, comparado con medios tradicionales como el periódico. Realizar un análisis de veracidad a cada una de estas noticias es una tarea imposible de realizar manualmente debido al esfuerzo humano requerido y al gran volumen de información que se produce cada hora, por lo que es necesario buscar medios de verificación automáticos que clasifiquen estos contenidos dentro de las primeras horas en que fueron emitidos. Este problema ha sido de gran interés para la comunidad académica donde se han creando diversos mecanismos para la detección de noticias falsas, principalmente basados en técnicas de deep learning y machine learning. No obstante, existen pocos trabajos específicamente diseñados para resolver la tarea de detección temprana, que utilicen tanto contenido como contexto para la clasificación. Por esta razón, en este trabajo proponemos un nuevo modelo de detección temprana de noticias falsas condicionado a las primeras etapas de la propagación. El modelo llamado Early Rumor Detection Model (ERDM), compuesto por una red Bi-GRU con un módulo de atención global, recibe en cada tiempo tanto características de la propagación de los mensajes (texto y tiempo) como información de los usuarios que participan en las conversaciones. Los resultados muestran que ERDM sobrepasa a los métodos de la literatura, tanto para escenario a 4 clases como binario, donde en este último escenario se consigue una mejora de 7% y 13.4% en los datasets Twitter 15 y Twitter 16 respectivamente. Además, ERDM supera los resultados del estado del arte en detección temprana obteniendo sobre 80% en accuracy en ambos datasets dentro de las primeras 4 horas de difusión de una noticia.
The massive use of social networks has allowed an explosive increase of fake news circulating on the web. The main reason is that this type of content can be created and published quickly and at zero cost, compared to traditional media such as newspapers. Performing a veracity analysis of each of these news items is an impossible task to perform manually due to the human e ort required and the large volume of information that is produced every hour, so it is necessary to find automatic methods of verification that classify these contents within the first hours after they were published. This problem has been of great interest to the academic community where several mechanisms for the detection of fake news have been created, mainly based on techniques of Ddeep learning and Machine learning. However, there are few works specifically designed to solve the early detection task, using both content and context for classification. For this reason, in this paper we propose a new model for early detection of fake news conditioned to the early stages of propagation. The model called Early Rumor Detection Model (ERDM), composed of a Bi-GRU network with a global attention module, receives at each time both features of the message propagation (text and time) and information of the users participating in the conversations. The results show that ERDM outperforms current detection methods, both for 4-class and binary scenarios, where in the last scenario an improvement of 7% and 13.4% is achieved in the Twitter 15 and Twitter 16 datasets, respectively. In addition, ERDM outperforms the state-of-the-art results in early detection, obtaining over 80% in accuracy in both datasets within the first 4 hours of news dissemination.
The massive use of social networks has allowed an explosive increase of fake news circulating on the web. The main reason is that this type of content can be created and published quickly and at zero cost, compared to traditional media such as newspapers. Performing a veracity analysis of each of these news items is an impossible task to perform manually due to the human e ort required and the large volume of information that is produced every hour, so it is necessary to find automatic methods of verification that classify these contents within the first hours after they were published. This problem has been of great interest to the academic community where several mechanisms for the detection of fake news have been created, mainly based on techniques of Ddeep learning and Machine learning. However, there are few works specifically designed to solve the early detection task, using both content and context for classification. For this reason, in this paper we propose a new model for early detection of fake news conditioned to the early stages of propagation. The model called Early Rumor Detection Model (ERDM), composed of a Bi-GRU network with a global attention module, receives at each time both features of the message propagation (text and time) and information of the users participating in the conversations. The results show that ERDM outperforms current detection methods, both for 4-class and binary scenarios, where in the last scenario an improvement of 7% and 13.4% is achieved in the Twitter 15 and Twitter 16 datasets, respectively. In addition, ERDM outperforms the state-of-the-art results in early detection, obtaining over 80% in accuracy in both datasets within the first 4 hours of news dissemination.
Description
Keywords
REDES SOCIALES , INTERNET , INFORMACION