Applying Machine Learning to Intelligent Assessment of Scientific Creativity Based on Scientific Knowledge Structure and Eye-Tracking Data

Scientific creativity plays an essential role in science education as an advanced cognitive ability that inspires students to solve scientific problems inventively. The cultivation of scientific creativity relies heavily on effective assessment. Typically, human raters manually score scientific creativity using the Consensual Assessment Technique (CAT), which is a labor-intensive, time-consuming, and error-prone process. The assessment procedure is susceptible to subjective biases stemming from cognitive prejudice, distractions, fatigue, and fondness, potentially compromising its reliability, consistency, and efficiency. Previous research has sought to mitigate these risks by automating the assessment through latent semantic analysis and artificial intelligence. In this study, we developed machine learning (ML) models based on a training dataset that included output labels from the Scientific Creativity Test (SCT) evaluated by human experts, along with input features derived from objectively measurable semantic network parameters (representing the scientific knowledge structure) and eye-tracking blink duration (indicating attention patterns during the SCT). Most models achieve over 90% accuracy in predicting the scientific creativity levels of new individuals outside the training set, with some models achieving perfect accuracy. The results indicate that the ML models effectively capture the underlying relationship between scientific knowledge, eye movements, and scientific creativity. These models enable the fairly objective prediction of scientific creativity levels based on semantic network parameters and blink durations during the SCT, eliminating the need for ongoing human scoring. Therefore, laborious and complex manual assessment methods typically used for SCT can be avoided. This new method improves the efficiency of scientific creativity assessment by automating processes, minimizing subjectivity, providing rapid feedback, and enabling large-scale evaluations, all while reducing evaluators' workloads.