Hybrid Smoothed Dirichlet-Family Regression Framework for Fake News Detection

Abstract

Standard multinomial models inadequately capture the sparsity, burstiness, and dependence present in highdimensional bag-of-words text. In this paper, we develop regression-based generative formulations of Smoothed Dirichlet (SD), Smoothed Generalized Dirichlet (SGD), and Smoothed Generalized Dirichlet Multinomial (SGDM) models for fake news detection from raw text counts. The models are expressed through distributional regression, where covariates are mapped by exponential links to positive distributional parameters. In the supervised setting, these links define class-specific probabilistic anchors that support both direct maximum a posteriori (MAP) classification and hybrid feature construction. We derive compact likelihood-based and Fisher-style features from the learned anchors and pass them to discriminative classifiers, preserving generative structure while relaxing rigid likelihood-based decision rules. Experiments on ISOT, PolitiFact, and GossipCop evaluate the framework across large balanced, small partially recovered, and highly imbalanced fake-news settings. Results demonstrate model behavior is strongly dataset-regime dependent: SGD benefits from larger balanced data, SD is more stable in small-data settings, and hybridization recovers useful signal when MAP classification collapses under imbalance. Overall, the proposed framework provides an interpretable probabilistic representation layer for sparse text counts and a structured way to combine smoothed generative modeling with discriminative classification.