Abstract
Click-through rate (CTR) prediction plays a pivotal role in modern online advertising ecosystems, directly influencing revenue optimization and user experience enhancement. Traditional CTR prediction models have primarily focused on homogeneous feature interactions, overlooking the inherently heterogeneous nature of advertising networks where multiple entity types interact through diverse relationships. This paper proposes a novel framework leveraging Graph Neural Networks (GNNs) to model these complex heterogeneous interactions for improved CTR prediction accuracy. Our approach integrates deep neural network architectures with pre-training strategies, attention-based feature interaction mechanisms, and adaptive activation functions to capture high-order dependencies and contextual information. The framework employs sparse neural network architectures with sampling-based pre-training, attentional factorization machines for dynamic feature weighting, and data-adaptive activation functions to enhance model expressiveness. Experimental validation on real-world advertising datasets demonstrates significant performance improvements over state-of-the-art baselines, achieving superior Area Under the Curve (AUC) and LogLoss metrics while maintaining computational efficiency suitable for industrial deployment.

This work is licensed under a Creative Commons Attribution 4.0 International License.
Copyright (c) 2025 Elena Marković, Jonas Frey (Author)