Abstract:

Electrical engineering students face the challenge of starting their studies with different pre-conceptions, especially in the first semesters, and so the challenges are manifold. In addition to subject-specific problems, language problems can also be a barrier when working on tasks such as sorting resistance networks according to resistance in ascending order. For this reason, we research how to support our students automated.

The target application is a VR environment where eye-tracking data can be collected. The application should automatically recognize students who need support when working on the tasks and offer support options accordingly. The basis for these support options is eye-tracking data. For this purpose, this eye-tracking data must be evaluated automatically.

In a course focusing on the fundamentals of electrical engineering, eye-tracking recordings were made of students working on the tasks. In each case, two tasks were recorded on three exercise sheets. The research question here is: Which metrics can be found solely based on eye-tracking data that can be used to differentiate between students who are completing the task correctly and students who are not? Building on this, which models are suitable for creating a forecast of success using these metrics?

As it has already been established that eye-tracking metrics that work for one task do not necessarily apply to another task, numerous eye-tracking metrics must be calculated for each task. The metrics are then examined for significance concerning the classification into "will be successful" or "will not be successful". For the task examined in this article, the eye-tracking metrics revisit count (RC), dwell (D), dwell rate (DR) and fixation proportion (FP) proved to be significant. In addition, dwell strings (DS) are used as a further metric. They represent the order in which the students visited the areas of interest (AOI). Thus, the students' approach is mapped in the DS.

To get any information out of a metric a model is necessary. In this case, we use different machine learning models, that fit the requirements like the decision tree (DT) that has a root, and every new branch of the tree needs a condition so that the leaves build the result and the support vector machine (SVM) that tries to maximize the mathematical distance between the classes. In addition, two models that are suited to model sequences (a hidden Markov model (HMM) and a (bidirectional) long short-term memory model ((B)LSTM) are used to predict the success of the students based on the DS. All of these models have in common that they need to be trained to get adapted to the data.

In addition, a hybrid model consisting of a (B)LSTM and a parallel convolutional neural network (CNN) that is specialized to extract features in the data is tested. The CNN is trained with the significant metrics that are also used to train the DT and the SVM. The (B)LSTM is trained with the DS. Finally, the outputs of the two subnetworks are combined. The hybrid model reaches an accuracy of up to 77,78 %.

In the article, the performance of the different models on one of the tasks will be compared. Also, the advantages and disadvantages of the different models that affect the target application mentioned above are discussed.

Keywords:

Eye-Tracking, Machine Learning, Learning System.