Semi-supervised learning involves incorporating information from unlabeled data in addition to a few labeled data, which directs the decision boundary to less dense regions. The main problem that we are trying to tackle is to find mental states of an individual, based on the pre-processed brain-wave signals. An electroencephalogram (EEG) time series dataset is used, which shows the local brain activity using electrodes attached to the brain. Input to the chosen algorithms is the pre-processed EEG data with 1532 sample points and nearly 2548 features that range from signal mean values to Fourier transformation amplitudes. In the original data, output is the predicted label (0 for relaxed, 1 for neutral and 2 for concentrating). This is converted into binary classes of relaxed and concentrating state. In addition to pure supervised methods, the baseline approach is chosen to be a pseudo-labeling method. As improvements on this approach, two different self-learning models (with Logistic Regression and Weighted Quadratic Discriminative Analysis) and Transductive Support Vector Machine (TSVM) models are applied.