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

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

摘要

Pinterest广告轻量级排名系统采用GPU服务的两塔模型，结合MMOE和DCN架构，显著提升了点击率预测的准确性。通过优化数据加载、模型代码和训练配置，训练效率大幅提高。离线损失减少5-10%，在线实验显示点击成本降低，点击率显著提升。这一改进为广告推荐系统带来了更高效、复杂的模型架构。