Redundancy Reduction in Composite Indicator Systems for Causal Graph Construction

Abstract

Indicator redundancy is common in capability assessments and, if uncontrolled, can bias results, reduce model performance, and distort causal graph learning. Using simulated datasets with known causal structures, this study examines how redundancy affects causal graph accuracy, complexity, stability, and model interpretability. We introduce MB-GAR, a Markov-blanket-guided genetic algorithm that removes redundant indicators while preserving information. Experiments show that MB-GAR markedly improves causal structure accuracy and predictive performance, outperforming correlation-based baselines by about 15% in redundancy-detection F1, demonstrating its effectiveness for building cleaner and more reliable evaluation indicator systems.