RaCMC: Residual-Aware Compensation Network with Multi-Granularity Constraints for Fake News Detection

TL;DR

RaCMC tackles multimodal fake news detection by introducing a multiscale residual-aware compensation (MRC) that fuses cross-modal features with a residual mechanism, and a multi-granularity constraints (MGC) module that sharpens discriminability at both the news and feature levels. A dominant feature fusion reasoning (DFR) module integrates unimodal and multimodal signals to reason about authenticity through consistency and inconsistency. The approach demonstrates state-of-the-art accuracy on Weibo17, Politifact, and GossipCop, underscoring robustness across datasets. This work advances cross-modal fusion by explicitly modeling interaction quality and distributional separability, with potential impact on reliable automated misinformation screening.

Abstract

Multimodal fake news detection aims to automatically identify real or fake news, thereby mitigating the adverse effects caused by such misinformation. Although prevailing approaches have demonstrated their effectiveness, challenges persist in cross-modal feature fusion and refinement for classification. To address this, we present a residual-aware compensation network with multi-granularity constraints (RaCMC) for fake news detection, that aims to sufficiently interact and fuse cross-modal features while amplifying the differences between real and fake news. First, a multiscale residual-aware compensation module is designed to interact and fuse features at different scales, and ensure both the consistency and exclusivity of feature interaction, thus acquiring high-quality features. Second, a multi-granularity constraints module is implemented to limit the distribution of both the news overall and the image-text pairs within the news, thus amplifying the differences between real and fake news at the news and feature levels. Finally, a dominant feature fusion reasoning module is developed to comprehensively evaluate news authenticity from the perspectives of both consistency and inconsistency. Experiments on three public datasets, including Weibo17, Politifact and GossipCop, reveal the superiority of the proposed method.

Figures (4)

• Figure 1: Comparison of the fusion and reasoning processes. (a) Previous methods exhibit inadequate cross-modal feature fusion and may misjudge real news with high similarity. (b) Our method refines cross-modal fusion with residual compensation and multi-scale information mining. Besides, it amplifies the differences between real and fake news in terms of news and feature levels.
• Figure 2: The network architecture of RaCMC. It consists of four components: (1) Feature Encoding: this module extracts features with different granularities utilizing a coarse-fine dual-grained encoder. (2) Multiscale Residual-aware Compensation: this module sufficiently interacts and fuses features from different modalities and sources. It has three branches, processing text (blue block), multimodal (yellow block), and image (orange block) features from top to bottom. The exact execution process is detailed in the image branch. (3) Multi-granularity Constraints: this module amplifies the differences between real and fake news at both the news and feature levels. (4) Dominant Feature Fusion Reasoning: this module evaluates news authenticity in terms of both consistency and inconsistency.
• Figure 3: t-SNE visualization of mined features on the test set of the Weibo17 dataset. Dots with the same color are within the same label.
• Figure 4: Ablation study of MRC, MGC and DFR under four metrics in three datasets.