A Hybrid Graph Attention and Ensemble Learning Framework for Predicting Consumer Engagement in Social Media Advertising

Abstract

The rapid growth of digital platforms has increased the complexity of predicting consumer engagement in social media advertising due to dynamic interactions among campaign features, audience behavior, and platform characteristics. Existing machine learning approaches primarily rely on feature-based representations and often fail to capture latent relational dependencies among advertising campaigns.

To address this limitation, this study proposes a novel hybrid framework that integrates Graph Attention Networks (GAT) with ensemble learning techniques, specifically Random Forest and XGBoost. A K-nearest neighbor-based graph is constructed to model inter-campaign relationships, enabling the extraction of high-quality relational embeddings through attention mechanisms.

These embeddings are further utilized within a stacking ensemble framework to improve prediction performance. Experimental results on a real-world dataset demonstrate that the proposed model achieves an accuracy of 87%, outperforming baseline models. The findings highlight the effectiveness of combining graph-based relational learning with ensemble modeling for robust and scalable consumer engagement prediction