Deep neural models for click-through rate prediction often exhibit a sharp decline in validation performance immediately after the first training epoch despite continued improvement in training loss. This instability restricts effective learning and limits model performance. This study analyzes this behavior using large-scale industrial datasets and evaluates practical mitigation strategies. While reducing the learning rate provides only incremental gains, controlling feature sparsity yields substantial improvements. Removing highly sparse features and aggregating infrequent feature values stabilizes training, extends useful learning time, and improves both offline evaluation metrics and online system performance.
Blogger's Review: This study reveals a crucial issue in CTR model training, particularly the impact of feature sparsity on model performance. Through empirical analysis and strategy adjustments, it demonstrates how to enhance model stability and effectiveness in real-world applications, making it a noteworthy reference.