The paper introduces a novel approach called Macro Graph Neural Networks (MacGNN) to address the computational complexity and sampling bias issues in Click-Through Rate (CTR) prediction for billion-scale recommender systems. Traditional Graph Neural Networks (GNNs) struggle with the vast number of neighbors in such systems, leading to high computational costs. To mitigate this, GNN-based CTR models often sample only a few hundred neighbors, which can result in significant sampling bias and an incomplete representation of user or item behavior. To tackle these challenges, the authors propose a Macro Recommendation Graph (MAG), which groups micro nodes (users and items) with similar behavior patterns into macro nodes. This reduces the number of neighbors from billions to hundreds, making it more efficient for GNNs to operate. MacGNN is designed to aggregate information at the macro level and update the embeddings of macro nodes. The model has been successfully deployed on Taobao's homepage feed, serving over one billion users. The paper includes extensive offline experiments on three public benchmark datasets and an industrial dataset, demonstrating that MacGNN outperforms twelve state-of-the-art CTR baselines while maintaining computational efficiency. Online A/B tests further confirm the superiority of MacGNN in real-world billion-scale recommender systems. The key contributions of the paper are the introduction of MAG and MacGNN, which significantly enhance the performance and efficiency of CTR prediction in large-scale recommender systems.The paper introduces a novel approach called Macro Graph Neural Networks (MacGNN) to address the computational complexity and sampling bias issues in Click-Through Rate (CTR) prediction for billion-scale recommender systems. Traditional Graph Neural Networks (GNNs) struggle with the vast number of neighbors in such systems, leading to high computational costs. To mitigate this, GNN-based CTR models often sample only a few hundred neighbors, which can result in significant sampling bias and an incomplete representation of user or item behavior. To tackle these challenges, the authors propose a Macro Recommendation Graph (MAG), which groups micro nodes (users and items) with similar behavior patterns into macro nodes. This reduces the number of neighbors from billions to hundreds, making it more efficient for GNNs to operate. MacGNN is designed to aggregate information at the macro level and update the embeddings of macro nodes. The model has been successfully deployed on Taobao's homepage feed, serving over one billion users. The paper includes extensive offline experiments on three public benchmark datasets and an industrial dataset, demonstrating that MacGNN outperforms twelve state-of-the-art CTR baselines while maintaining computational efficiency. Online A/B tests further confirm the superiority of MacGNN in real-world billion-scale recommender systems. The key contributions of the paper are the introduction of MAG and MacGNN, which significantly enhance the performance and efficiency of CTR prediction in large-scale recommender systems.