Scanpath Analysis of Programming Comprehension: A MultiMatch and Clustering-Based Approach

In programming studies, inferring participants’ cognitive processing from eye movement behavior provides valuable insights into Performance Level (PL) and problem-solving strategies. This paper explores scanpath variation in eye-tracking data from 60 under-graduate students completing six multiple-choice programming tasks (MCPTs), each with multiple correct alternatives. Responses were categorized as Correct (C) or Partially Correct (P) and further grouped into High and Low Performance levels (PLs). Visual scanpaths were analyzed using the MultiMatch algorithm across five dimensions: Length, Position, Duration, Shape, and Direction. Pairwise similarity scores were computed to reveal patterns of scanpath behavior, and K-means clustering with standard validity indices (Silhouette Score, Inertia, Davies–Bouldin Index) was applied to evaluate group-level differences. Results showed that Length and Position consistently reflected task structure, while Duration, Shape, and especially Direction captured meaningful variation in attention allocation, scanpath geometry, and sequencing. Ultimately, this study demonstrates that how students look, as reflected in temporal and directional scanpath features, is more informative for assessing cognitive strategies than where they look, offering practical opportunities to enhance adaptive learning technologies and real-time pedagogical support.