A utilização de sistemas de recomendação no contexto da desinformação

Abstract

The misinformation “pandemic”, facilitated by social media and mobile messaging platforms, has gone from being a minor nuisance to severely impacting law and order through the deliberate and large-scale manipulation of the general public’s sentiments. Recommender systems are an essential set of techniques and software tools capable of acting on the problem of content filtering and information retrieval in an overwhelming scenario of choice possibilities that presents the technological world today. One platform with a large data flow in the Brazilian scenario is the WhatsApp application, which has been the target of agents sharing disinformation content. With the present work, we intend to experiment to discover how recommender systems could act to recommend messages of a similar character concerning their factual content. We will demonstrate how the employed techniques could represent a recommender system, presenting their architectures, similarity methods, lists of recommendations and performance evaluation mechanisms. We will also present how the data under study, regarding messages collected from WhatsApp concerning the COVID-19 pandemic, behaved with the proposed architecture, as well as their results regarding the performance of recommendations, in order to understand how performance metrics such as mean reciprocal rank and mean average precision characterize the relevance brought by the results when using a simple criterion of similarity - cosine similarity - in purely textual contents. Finally, the ability of the recommendation system itself to be used as part of a classifier for messages that do not have a assigned label of disinformation presence will be evaluated, since the classifier uses a naive approach to identify the presence (or absence) of disinformation content.