Exploring Critical Eye-Tracking Metrics for Identifying Cognitive Strategies in Raven’s Advanced Progressive Matrices: A Data-Driven Perspective

To explore the efficacy of various features in predicting cognitive strategy usage, this study employed a recursive feature elimination approach in conjunction with a random forest algorithm to identify the most effective predictors. In addition to item response accuracy (RA) and response time (RT), five key eye-tracking metrics were considered: proportional time on matrix (PTM), latency to first toggle (LFT), rate of latency to first toggle (RLT), number of toggles (NOT), and rate of toggling (ROT). The results indicated that PTM, RLT, and LFT are the three most critical features for predicting cognitive strategy usage, with PTM being the most important, followed by RLT and LFT. Clustering analysis of the optimal feature subset (PTM, RLT, and LFT) further validated the efficacy of these eye-tracking metrics in predicting cognitive strategy usage.