Combining compact news representations generated using DistilBERT and topological features to classify fake news

Abstract

Fake news (FN) have affected people’s lives in unimaginable ways. The automatic classification of FN is a vital tool to prevent their dissemination and support fact-checking. Related work has shown that FN spread faster, deeper, and more broadly than the truth on social media. Besides, deep learning has produced state-of-the-art solutions in this field, mainly based on textual attributes. In this paper, we propose initial experiments to combine compact representations of the textual news properties generated using DistilBERT, with topological metrics extracted from the social propagation network. Using a dataset related to politics and five distinct classification algorithms, our results are encouraging. Regarding the textual attributes, we reached results comparable to state-of-the-art solutions using only the news title and contents, which is useful for FN early detection. The topological attributes were not as effective, but the promising results encourage the investigation of alternative architectures for their combination

Assessing the combination of DistilBERT news representations and difusion topological features to classify fake news

“…Our results were promising, showing that the classification of FN based only on the title and content of the news achieves results close to the state-of-theart Liao et al 2021], which also consider the text of the propagation posts; -evaluation of the contribution of topological features of propagation networks as representative attributes of social engagement, previously restricted to identifying communication vehicles that propagate false news [Pierri et al 2020]. We improved our previous results [Sáenz et al 2020] by considering additional topological features. -an encompassing experimental setting to assess all components of the proposed approach.…”

“…The metrics to be calculated, demonstrated in [Pierri et al 2020] to have managed to reach several cases in diffusion networks, such as when users within the network form groups, networks where there is no mono-directionality in the diffusion of news or networks where there is a single user who distributes the news among all others (broadcast). The metrics adopted in [Pierri et al 2020], and experimented in our original work [Sáenz et al 2020 In this article, we included four additional metrics that characterize the network as a whole and that were used in some other works [Zhou and Zafarani 2019;Shu et al 2020], to determine if they can contribute to the improvement of the results: To calculate these topological attributes, we first uploaded the news with their tweets and retweets to the graph-oriented database neo4j 6 . There, each news, tweets and retweets were represented as nodes of an heterogeneous network.…”

“…This second set of experiments aims to analyze the discriminating power of topological metrics for fake news classification. For this purpose, we carried out several experiments, namely: a) assessment of the proper representation for topological metrics for classification purposes; b) an evaluation of the predictive power of the topological metrics; and c) a performance comparison between our previous work [Sáenz et al 2020] and the models constructed with a refined set of topological metrics. These experiments consider the news for which we were able to collect the respective diffusion network, and thus the dataset includes only 659 news (Table I).…”

“…Thus, all results reported in the remaining of this section were obtained using the topological metrics represented as z-scores. In our previous work [Sáenz et al 2020], we used the topological metrics originally proposed in [Pierri et al 2020], which we consider our baseline for this set of experiments. In an attempt to improve these previous results, we considered four additional topological features.…”

“…We assessed the contribution of each type of feature separately and their combination. This article is an extension of our previously presented work [Sáenz et al 2020]. We have significantly evolved it by: a) extending the set of topological features used for classification and assessing their contribution for FN classification; b) extending the experimental settings, and analyzing in more detail the performance results; c) leveraging stacking ensembles to improve the classification of fake news; and d) updating the theoretical background and related work.…”