Emotion recognition based on EEG signals has become a prospective research area because of its potential application. The features used for classifier input play an important role in the classification results, while there is no exact method for defining the best feature for a classifier. In this paper, we use nine types of time frequency domains as features. We also use some feature selection methods to select the best feature for classification. We compare the results of classifications from each method of feature selection. This method is implemented using a DEAP (Dataset for emotion analysis using psychological signal) dataset. The results of the experiment demonstrate that the time-frequency domain feature extraction shows the best performance when using PCA feature selection and the k-NN classifier with a 60.68% recognition rate. This experiment also shows that there is no significant result between k-NN and weighted k-NN classification for each feature selection. In this case, we can conclude that the DEAP data needs a more reliable feature selection method.