Investigating Visual Attention and Behavioral Patterns of High- and Low-Performers in a Dual-Scaffold Programming Environment

Description

Programming learning environments increasingly provide scaffolding, ranging from static prompts to interactive artificial intelligence (AI)-based support. When scaffolds are available, learners must decide which support to attend to and how to integrate information during problem solving. However, little is known about how learners with programming performance levels coordinate resources. This study employed eye-tracking, lag sequential analysis (LSA), and retrospective think-aloud data to examine high- and low-performers’ learning processes in a dual-scaffold programming environment. Low-performers exhibited more frequent fixations across scaffold-related areas than high-performers, with a significant main effect of performance group on fixation count (F = 6.854, p = .011, ηp² = .081). Scanpath analysis and LSA showed that high-performers demonstrated coherent task-oriented patterns linking code construction, debugging, and AI-supported checking or explanation, whereas low-performers used scaffolds for task confirmation and code acquisition. This study provides implications for designing adaptive AI-supported programming environments and instructional strategies for novice programmers.

Steps to reproduce

The dataset was collected from a study investigating students’ visual attention and behavioral patterns in a dual-scaffold programming environment. The study involved 100 seventh-grade students from a junior high school in China. After obtaining informed consent and ethical approval, students completed an introductory Python programming task in an environment providing both a static prompt scaffold and an interactive AI scaffold. The task covered fundamental Python concepts, including variable assignment, output statements, and sequential execution. The environment was implemented in Visual Studio Code (VS Code), with Tongyi Lingma providing AI support. Students could actively request code explanations, debugging suggestions, and code checking, while automatic AI-assisted functions were disabled. Eye-tracking data were collected using a Tobii Pro X-60 eye tracker (60 Hz) and Tobii Pro Lab software after gaze calibration. Five areas of interest (Question, Hints, AI, Code, and Data) were defined, and fixation count (FC), total fixation duration (TFD), and average fixation duration (AFD) were extracted. Screen recordings were collected using EV screen recording software to capture programming behaviors. These behaviors were coded into ten categories: understanding task (UT), referring to the static prompt scaffold (RSP), coding in Python (CP), debugging in Python (DP), running codes (RC), asking AI new questions (AANQ), asking AI to explain codes (AAEC), asking AI to check codes (AACC), pasting Python codes (PPC), and copying and pasting codes (CPC). The coded sequences were analyzed using lag sequential analysis. Students’ programming knowledge was assessed through a test covering fundamental Python concepts. Two teachers independently scored the test, and the average score was used for grouping. Using an extreme-group approach, the top 40% of students were assigned to the high-performer group (n = 40), the bottom 40% to the low-performer group (n = 40), and the middle 20% were excluded. Students participated in retrospective think-aloud interviews; these data were used only for supplementary interpretation and were not systematically coded or statistically analyzed, so they were not included in the released dataset. The released dataset contains anonymized programming knowledge test scores, eye-tracking measures, scanpath sequences, and behavioral coding files for lag sequential analysis.

Institutions

• Shaanxi Normal University

  Shaanxi, Xi'an

Categories

Licence