A hybrid contrastive graph neural network framework for fake news and hate speech detection using content and user interaction signals

Detecting fake discussions and aggressive speech on social platforms remains a complex task due to evolving user dynamics and the minimal overlap between normal and harmful content. Traditional detection methods lag in classifying these contents as they focus only on textual features or user behavior. This leads to limited performance specially in sparse interaction environments. To overcome these limitations a hybrid graph-based classification model is proposed in this research work which incorporates semantic, behavioral, and contextual data into a combined architecture. The proposed system models content and users as nodes, with interaction types such as replies, mentions, and shares to create a forming weighted edges based on temporal and behavioral factors. Also, a contrastive strategy is incorporated to differentiate the aligned and conflicting user-content associations. The textual representation in the comments is captured using a semantic encoder and the relational dependencies are modeled through graph attention layers which further enhanced with metadata like credibility scores and user activity. Experimental evaluations using the FakeNewsNet dataset confirm the proposed model superior performance through the attained 97.12% accuracy which is superior over conventional methods like GRU-MCAF (94.82%) and Attention-LSTM (93.57%). The proposed model also recorded the highest F1-score of 96.43%, precision of 96.72%, and recall of 96.15% which indicates the model consistent classification across labels over conventional methods.