With numerous real-world applications, eye tracking is a versatile tool for a wide range of research areas. Recently, this technology has found its way into classrooms through pedagogical research. By measuring a person's gaze over time, conclusions can be drawn about their cognitive processes, strategies, and particular difficulties during tasks. For example, eye tracking enables educators and researchers to model learners by determining their gaze patterns. Furthermore, eye tracking can be used directly for teaching by extracting and learning from gaze behaviour of experts in particular fields.

However, to exploit the full potential of eye tracking, it is necessary to classify explicit eye movements, such as fixations and saccades, from the data obtained. In current eye tracking software, these gaze movements are measured using classical algorithms with thresholds that have to be set manually by educators or researchers. However, with noisy data, these calculations are prone to mislabelling, which can lead to incorrect conclusions and therefore reduced quality of insights and learning. Deep learning approaches can eliminate the need for manual threshold setting and counteract the influence of data noise.

This paper investigates the performance of a deep learning model for eye movement classification in regard to its generalisability to a new and fully hand-labelled test data set. To achieve this, we use a model that has already proven itself in other literature - an open-source 1-dimensional convolutional neural network with bidirectional long short-term memory trained on the GazeCom data set. Hereby, the original training data was recorded at a sampling frequency of 250 Hz with the SR Research EyeLink II. For recording the test data set, we utilise the Tobii Pro Spectrum eye tracker, which cannot operate at the sampling frequency of 250 Hz. Therefore, the aim of this work is to enhance the applicability of the original model to also label data sets recorded with the Tobii Pro Spectrum at frequencies of 300 Hz and 600 Hz. Thus, several models based on the originally published model, though with different parameters, are trained on the original data set.

In conclusion, this study shows that modifying and training the original model with different parameters than in the original publication allows it to perform satisfactorily, even though it was not initially able to label the Tobii Pro Spectrum data set. This result, together with the newly trained deep learning model, enables educators and researchers working with the Tobii Pro Spectrum eye tracker to classify eye movements using a modern approach. This will allow for more accurate identification of eye movements and gaze patterns, thereby ensuring more reliable results for the classroom of the future.

Keywords: Eye tracking, fixation, saccade, classification, deep learning.