This paper introduces Macro Graph Neural Networks (MacGNN) for billion-scale online recommender systems. The main challenge is the computational complexity of aggregating billions of neighbors in Graph Neural Networks (GNNs) for CTR prediction. To address this, the authors propose a Macro Recommendation Graph (MAG) that groups user and item nodes with similar behavior patterns into macro nodes, reducing the number of neighbors from billions to hundreds. This allows for efficient computation and accurate CTR prediction. MacGNN is designed to aggregate information on a macro level, updating macro-node embeddings to enable efficient online CTR prediction. The model has been deployed on Alibaba's Taobao platform, serving over one billion users. Extensive offline experiments on three public benchmark datasets and an industrial dataset show that MacGNN significantly outperforms twelve CTR baselines while maintaining computational efficiency. Online A/B tests confirm MacGNN's superiority in billion-scale recommender systems. The key contributions of this work include: (1) creating a customized macro recommendation graph that reduces neighbor size from billions to hundreds, (2) proposing a novel macro-scale recommendation paradigm known as the Macro Graph Neural Network (MacGNN), and (3) demonstrating that MacGNN has been serving a major shopping platform for two months, offering recommendations to over one billion users. The authors also address the challenges of sampling bias, unfitted users/items sampling, and ambiguous neighbor counts in traditional CTR models. By grouping nodes into macro nodes, the model eliminates the need for sampling and reduces computational complexity. The macro graph paradigm allows for more accurate and efficient CTR prediction by capturing the behavior patterns of users and items at a macro level. The model's effectiveness is validated through extensive experiments and real-world deployment.This paper introduces Macro Graph Neural Networks (MacGNN) for billion-scale online recommender systems. The main challenge is the computational complexity of aggregating billions of neighbors in Graph Neural Networks (GNNs) for CTR prediction. To address this, the authors propose a Macro Recommendation Graph (MAG) that groups user and item nodes with similar behavior patterns into macro nodes, reducing the number of neighbors from billions to hundreds. This allows for efficient computation and accurate CTR prediction. MacGNN is designed to aggregate information on a macro level, updating macro-node embeddings to enable efficient online CTR prediction. The model has been deployed on Alibaba's Taobao platform, serving over one billion users. Extensive offline experiments on three public benchmark datasets and an industrial dataset show that MacGNN significantly outperforms twelve CTR baselines while maintaining computational efficiency. Online A/B tests confirm MacGNN's superiority in billion-scale recommender systems. The key contributions of this work include: (1) creating a customized macro recommendation graph that reduces neighbor size from billions to hundreds, (2) proposing a novel macro-scale recommendation paradigm known as the Macro Graph Neural Network (MacGNN), and (3) demonstrating that MacGNN has been serving a major shopping platform for two months, offering recommendations to over one billion users. The authors also address the challenges of sampling bias, unfitted users/items sampling, and ambiguous neighbor counts in traditional CTR models. By grouping nodes into macro nodes, the model eliminates the need for sampling and reduces computational complexity. The macro graph paradigm allows for more accurate and efficient CTR prediction by capturing the behavior patterns of users and items at a macro level. The model's effectiveness is validated through extensive experiments and real-world deployment.