iQIYI's RSLIME: A Novel Feature Importance Analysis Method for Video Recommendation Systems

This paper introduces iQIYI's small video recommendation system and proposes a new feature importance analysis method called RSLIME (Recommendation System Boosted Local Interpretable Model-Agnostic Explanations Method). The system addresses four major challenges in UGC video recommendation: freshness, cold start, diversity, and interest shift.

The recommendation system follows a three-stage architecture: user profiling, recall, and ranking. The ranking module is particularly complex, integrating multiple models including DNN, GBDT, and FM. To address the interpretability challenge of this complex system, iQIYI developed RSLIME, which provides feature importance analysis without being affected by the underlying model architecture.

RSLIME has three key characteristics: it can generate feature importance estimates for individual input samples regardless of the ranking module's architecture, it can perform comprehensive analysis by combining predictions from multiple samples, and it can efficiently analyze the impact of sparse features to guide model optimization and feature selection.

The paper's main contributions include: detailed explanation of iQIYI's small video ranking module architecture, proposal of an interpretable recommendation system analysis method (RSLIME), and comprehensive experimental validation of RSLIME's effectiveness on iQIYI's ranking module.

The technical implementation uses an extended Deep&Wide structure that integrates GBDT, FM, and DNN. The system employs a sigmoid layer as a connecting layer that takes inputs from both the DNN's final hidden layer and FM's output. The paper also describes the implementation details of RSLIME for both single-case and multi-case analysis using MapReduce for distributed processing.

Experimental results demonstrate RSLIME's effectiveness in feature importance analysis, showing its ability to provide reliable and stable feature importance estimates and its correlation with AUC Check. The method proves valuable for understanding and optimizing recommendation system performance.