Handling Delayed Labeling of EEG Data Stream Using Semi-Supervised Label Propagation

Research interest in data stream classification is increasing through the development of adaptive machine learning techniques. These techniques involve continuously adjusting the classification model in response to changes in the data distribution. Most of these techniques assume instance labeling for the classes to perform the model adapting process, and this assumption is rare with actual data. This work proposes using a semi-supervised label propagation technique to infer many delayed labels (considered missing values) from limited known values in a data stream. The work's implementation included using two imbalanced EEG datasets, CUB-MIT Scalp and Siena Scalp datasets, to evaluate the proposed method with various values for missing ratios. The results showed the proposed method's ability to recover all the negative class values in both datasets with a missing percentage reaching 70%. Due to the rare positive class, the recovery of its value decreased with more than 30% missing ratio.