Large-Scale User Visits Understanding and Forecasting with Deep Spatial-Temporal Tensor Factorization Framework

This article discusses the challenges and solutions in advertising inventory prediction, focusing on a deep spatial-temporal tensor factorization framework. The framework addresses three main challenges: handling large-scale data and attributes, understanding complex user attributes, and considering both short and long-term changes. The model uses attention embedding mechanisms and multi-task training to improve prediction accuracy while reducing computational resources. Tested on real-world data from Tencent Video, the model showed significant improvements over traditional methods like ARIMA and CNN-based models, with a 15.6% reduction in standard deviation compared to ARIMA. The model has been successfully deployed in Tencent's online advertising system, leading to increased revenue and better user targeting.