Rslime: An Efficient Feature Importance Analysis Approach For Industrial Recommendation Systems

iQIYI represents one of the largest scale and most sophisticated video recommendation systems in China. In this paper, we focus on the model architecture as well as feature interpretation of our short video recommendation system. We firstly describe the short video recommendation system of iQIYI at a high level, which follows a three-stage information retrieval dichotomy. A Recommendation System Boosted Local Interpretable Model-Agnostic Explanations (RSLIME) is then proposed for real-time feature interpretation and importance evaluation. Lastly, comprehensive online and offline experiments are conducted to demonstrate the effectiveness of RSLIME and prove its unique advantages in the task of feature importance analysis and feature selection on large scale industrial recommendation systems.