EVALUATING ENSEMBLING METHODS FOR EYE STATE CLASSIFICATION

Abstract

Eye state tracking is one of the well-researched problems. EEG is a frequently used sensor modality to capture the state of human cognition. This is also a widely studied problem in the field of Computer Vision. Tracking eye movements with the help of a camera may be obtrusive (camera placement may only sometimes be conducive in each environment). In such a scenario, approaches like EEG measurement are preferred to tracking eye movements using a camera. Since this is a non-intrusive approach (and painless), this is preferable. In this work, we explore a supervised classification approach to identifying eye states (open or closed). We compare using widely used approaches such as LightGBM, Random Forests (Decision Trees), and XGBoost. These three classification approaches use ensemble-based techniques to aggregate decisions (sequential tree building, bagging, and boosting). These ensemble methods are preferred over other classifiers as they aggregate decisions over several classifiers and improve generalization. There are over 14,000 samples in this dataset (EEG Eye State Dataset). This is relatively small. Hence, our approaches use simple algorithms (as Deep learning-based algorithms usually require large training datasets). We use k-fold cross-validation to evaluate our results over several folds. This way, we ensure that performance is generalized and not dependent on a specifically chosen train or test set.