The use of eye-tracking in immersive Virtual Reality (iVR) is becoming an important tool for improving the learning outcomes. Nevertheless, the best Machine Learning (ML) technologies for the exploitation of eye-tracking data is yet unclear. Actually, one of the main drawbacks of some ML technologies, such as classifiers, is the scarce labeled data for training models, being the process of data annotation time-consuming and expensive. This paper presents a complete experimentation where different ML algorithms were tested, both supervised and semi-supervised, for trying to identify the stressors/distractors present in iVR learning experiences simulating the operation of a bridge crane. Results shown that the use of semi-supervised techniques can improve the performance of the Machine Learning methods making possible the identification of stressful situations in iVR environments. The use of semi-supervised learning techniques makes possible training ML algorithms without the need of great amount of labeled data which makes the data exploitation cheaper and easier.
