H-PLE: Algorithmic modeling of recommendation systems based on hierarchical PLE

In this study, we propose H-PLE (Hierarchical Progressive Layered Extraction), a novel multi-task recommendation framework designed to alleviate negative transfer and enhance performance under extreme abel imbalance. By introducing a hierarchical expert structure combined with improved gating and feature interaction modules, H-PLE effectively balances shared and task-specific knowledge extraction. We evaluate the model on large-scale real-world datasets and compare it against established baselines such as MMoE and PLE. The experimental analysis demonstrates that H-PLE consistently achieves stronger ranking performance, particularly in sparse-label tasks, while maintaining stable optimization behavior across multiple runs. Moreover, we reveal a practical gap between high ranking metrics and low decision-level metrics under default thresholding, and provide solutions through probability calibration, validation-based threshold selection, and top-K or rate-controlled strategies. These findings not only highlight the methodological advantages of H-PLE but also contribute actionable insights for bridging the gap between offline evaluation and online deployment in large-scale recommendation systems.