A Moving Target: Detecting Concept Drift in Brazilian Portuguese Fake News

Download PDF
Open Webpage

Manuela Guedes Wanderley, Lucca Baptista Silva Ferraz, Tiago Agostinho de Almeida, Renato Moraes Silva

Published: 2025, Last Modified: 20 Jan 2026STIL 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Resumo Static fake news detectors trained on offline data degrade over time due to concept drift, an understudied phenomenon outside English. This paper presents the first large-scale analysis of concept drift in Brazilian Portuguese fake news. Combining statistical two-sample tests, semantic similarity, and nonparametric change point detection, we quantify the presence and impact of drift, providing explainability by identifying points in time where shifts occur. Our results reveal significant shifts in topical and semantic patterns, demonstrating that model performance can degrade considerably when trained on older data. These findings prove the critical need for adaptive, time-aware methods, and the curation of temporally diverse datasets to build robust defenses against online misinformation in Brazilian Portuguese. The source code for our experiments is publicly available at https://github.com/GDSMN/STIL2025_conceptdrift.