Position-Aware Metric Normalization: A Hierarchical Framework for Context-Sensitive Evaluation of Search and Recommendation Systems

Abstract

Position bias and creative format effects significantly impact the evaluation accuracy of modern search and recommendation systems, yet traditional metrics often fail to account for these complex interaction patterns. The article presents a comprehensive evaluation framework that implements sophisticated normalization techniques to address these challenges. The framework introduces a hierarchical correction model that accounts for vertical and horizontal position bias while simultaneously considering viewport visibility patterns and scrolls depth distribution. The system implements a creative-aware correction model that captures format-specific engagement baselines, cross-format interaction effects, and temporal attention patterns. By integrating fine-grained viewport tracking and precise interaction event collection, the framework enables more accurate performance assessment by normalizing metrics across multiple dimensions. Experimental results demonstrate that the approach significantly improves evaluation accuracy compared to traditional metrics, leading to more informed optimization decisions in search and recommendation systems. The framework's adaptive nature, powered by automated learning of interaction patterns and dynamic adjustment of normalization factors, makes it particularly suitable for contemporary applications where user behavior patterns continuously evolve.