Abstract:

Creativity is one of the so-called 21st-century skills and has been included in the official curricula of many countries (Patston, et al., 2021). For the assessment of creativity, most creativity tools involve graphical solutions (images) evaluated by experts. The scoring with the engagement of human raters requires time and effort (Wang, et al., 2008), and the automated scoring methods are insensitive to the content (Harbinson et al., 2014). As an alternative to both expert judgments and fully automatic scoring, machine learning (ML) demonstrated a great contribution to the field of educational assessment.

The present study investigated the possibilities of assessing student creativity based on ML for image recognition.

As a tool for measuring creativity in primary school, we developed a computer-based interactive environment that allows students to create unique images by choosing elements from a fixed set. Moreover, the two versions of the tool were developed. The first one has no constraints on the number and type of elements. Several constraints were implemented into the second version to make the interpretation of the work products more straightforward. To describe images in terms of elaboration and originality (subcomponents of creativity) indicators were developed for each image (Uglanova et al., 2020). Indicators are predefined key characteristics of images that can be transformed into quantitative data.

The overall sample size consisted of 3000 Russian primary school students (10-11 years old). Each student created three images that were analyzed as independent student work products.

The data analysis was organized in two steps. In the first step the latent class analysis (LCA) was applied to classify images into classes by their creativity level using data obtained from student actions in the computer-based environment (indicators). This step allowed us to obtain data labels to train the neural network without experts’ involvement.

In the second step, we apply ML methods. Of all ML algorithms, Deep Learning is the most effective (Schmidhuber et al., 2015). Deep learning is a collection of ML based on learning representations, rather than specialized algorithms for specific tasks. The "depth" of learning is due to the multilayered artificial neural networks. In the research, pretrained convolutional neural network Densenet was used, which was further trained on images.

The results from the LCA indicated a two-class solution as preferable for both creativity subcomponents in both instrument versions. Images in the first class were considered more creative than images in the second class. The accuracy performance of neural networks is high for both originality and elaboration (88% and 87,6%) for the first version and slightly higher for the second version.

The study addresses the research on the use of ML in the field of educational assessment. Particularly, we focus on the application of ML to image recognition with the purpose of creativity assessment. Creativity assessment is a challenge for educators and psychometricians. To provide the assessment system with valid and immediate results, the two-step methodology was presented. The results demonstrated the high accuracy in network predictions for identifying the level of creativity based on the results of LCA analysis.

The study postulates large-scale prospects for ML to assess complex educational and psychological characteristics.

Keywords:

Creativity, image analysis, neural networks, educational assessment, psychometrics, machine learning.