Few-shot learning with distribution calibration for event-level rumor detection

• Published in: Neurocomputing (Amsterdam) 2025-02, Vol.618, p.129034, Article 129034
• Main Authors: Ran, Hongyan, Jia, Caiyan, Li, Xiaohong, Zhang, Zhichang
• Format: Article
• Language: English
• Subjects: Calibrate distribution; Event-invariant; Few-shot setting; Newly emergent events; Rumor detection
• ISSN: 0925-2312
• DOI: 10.1016/j.neucom.2024.129034

With the rapid evolution of social media, rumors travel at unprecedented speeds. Automatic recognition of rumors is important for making users receive truthful information and maintaining social harmony. Recently, deep learning models have demonstrated strong rumor detection ability by capturing semantics of rumors, learning rumor propagation patterns and integrating users’ properties, etc. However, most existing rumor detection models perform poorly on unseen events because they are powerful at capturing event-specific features in seen data, which cannot be transferred to newly emergent events due to distribution differences between events. Therefore, in this study, we propose a novel model named E-Rumor in a few-shot setting to learn and transfer event-invariant features from historic events with sufficient samples to new events with only a few examples. The model first calculates the rumor class distributions of historic rumors, and then calibrates the class distributions of a new event with the old ones. Furthermore, an adequate number of samples are generated from the calibrated distributions to expand the training set for a new-event rumor classifier. The proposed model can be paired with any rumor detection classifier as a feature extractor without extra parameters. Empirical studies have shown that a simple MLP (Multilayer Perceptron) trained on the samples generated from the calibrated distributions can outperform the state-of-the-art baseline models on two benchmark rumor datasets PHEME5, PHEME9 and one topic-level dataset T-Twitter, which is generated by topic extraction and cross-data manner, with more than 15.15%, 4.7% and 7.38% accuracy improvements, respectively.