Transfer Learning-Based Multi-Dimensional PE Assessment Framework

Abstract

Evaluating the performance of students during PE lessons is a challenging task that entails many aspects. The assessment does not only cover the students' fitness levels but also includes their social and psychological traits. However, such criteria are often ignored by automatic assessment systems as they employ single-model and single-dimensional methods. This paper introduces a data-driven evaluation framework for students' performance which takes into account all three correlated learning dimensions: physical, psychological, and social through the use of a predictive pipeline that employs a threephase transfer learning method. Our framework is based on a weighted ensemble which combines three machine learning algorithms: backpropagation neural network (PyTorch), random forest, and XGBoost regressor. Phase 1 deals with pretraining. In Phase 2, we fine-tune each model separately using the data from 613 real-world instances of the target domain so that the predictions would be consistent with the specifics of the environment. In Phase 3, softmax-based weighting dynamically suppresses unstable models in order to minimize the contribution of the models sensitive to covariate shift and shift the emphasis of the ensemble onto the most reliable cross-domain models. Our approach employs an effective two-tiered adaptive quiz engine to identify psychosocial predictors by means of psychological and social scenario quizzes in Tier 2 and the questionnaire in Tier 1. The post-hoc interpretation of the models via PFI indicates that goal orientation, motivation, and coordination are the main predictors of student performance. On the test set containing 613 real-world instances, our approach achieves an $R^{2}$ score of 0.9693 and MAE = 1.4114 points out of 100.