A Multi-Granular Adaptive Fusion Network (MGAF-Net) for Fake News Detection
A Multi-Granular Adaptive Fusion Network (MGAF-Net) for Fake News Detection
Mina Rajabi1
1) Department of Computer Engineering, Golestan University, Gorgan, Iran, Email:
IOI Code
:
AAHD-ABKEDE
Publication :
National Conference on Information Technology, Nanotechnology, Artificial Intelligence and Technological Futures Studies(itnaf.ir)
Abstract :
The proliferation of fake news on social media platforms poses a critical threat to public discourse
and democratic processes, yet existing detection approaches often rely on single-granularity features
or static ensemble methods that fail to capture the complex, multi-faceted nature of misinformation.
While deep learning models excel at sequential pattern recognition and traditional machine learning
captures handcrafted linguistic indicators effectively, few studies systematically evaluate whether
optimized fusion of heterogeneous representations provides measurable benefits over simpler
combination strategies under controlled conditions. This paper proposes MGAF-Net, a MultiGranular Adaptive Fusion Network that integrates TF-IDF, lexical, syntactic, and semantic features
through parallel XGBoost, BiLSTM-Attention, and MLP branches, with grid-search optimization
determining combination weights [α, β, γ]. Experiments on the ISOT Fake News Dataset (n=3,000,
5 random seeds) demonstrate that MGAF-Net achieves 99.02% ± 0.27% accuracy, significantly
outperforming Logistic Regression (96.71%, p=0.0038) and BiLSTM-only (81.29%, p=0.0001)
baselines. However, our most important empirical finding reveals that grid-search-optimized
weighted fusion does not improve performance over fixed equal-weight fusion (both 99.02%),
despite correctly identifying XGBoost as the dominant component (α≈0.68). The marginal 0.18%
gain over XGBoost alone (98.84%) and the lack of significant improvement from weight
optimization (p=0.0993 versus XGBoost) indicate that multi-granular feature engineering
contributes more to performance than adaptive fusion mechanisms for this dataset. These results
provide valuable insights into when ensemble weight optimization provides benefit versus when
strong baselines limit potential gains.
Keywords : fake news detection multi granular features grid search optimized fusion bilstm xgboost ensemble learning isot dataset