GPU-Serving Two-Tower Models for Lightweight Ads Engagement Prediction

Summary

The article presents a significant advancement in Pinterest's ads recommendation system through the introduction of a GPU-serving two-tower model for lightweight ranking. This model architecture combines Multi-gate Mixture-of-Experts (MMOE) with Deep & Cross Networks (DCN), optimizing both model performance and serving latency. The transition from CPU to GPU serving has led to a notable reduction in offline loss for click-through rate (CTR) prediction, achieving a 5-10% improvement. The article also highlights various enhancements in training efficiency, including dataloader optimizations and model code efficiency improvements, which collectively contribute to faster training times and better resource utilization. The evaluation results indicate substantial gains in both offline and online metrics, underscoring the effectiveness of the new architecture in scaling Pinterest's recommender systems.

Key Learnings

• 1The transition to GPU-serving for the two-tower model significantly enhances the efficiency of ad engagement prediction by reducing latency while maintaining performance.
• 2Incorporating MMOE with DCN allows for better handling of multi-domain multi-task challenges without relying on domain-specific modules.
• 3Optimizations such as GPU prefetching and BF16 precision training can drastically improve training times and resource utilization.
• 4Segregating ad scenarios during training leads to improved model performance and faster iteration speeds, demonstrating the importance of tailored data handling.
• 5Evaluation metrics like cost-per-click (CPC) and click-through rate (CTR) are critical for assessing the success of the model in real-world applications.

Who Should Read This

Senior Machine Learning Engineers focusing on optimizing ad recommendation systems and improving model serving efficiency.

Test Your Knowledge

Topics