Of late, emotion detection from brain signal has become a topic of research. Various machine learning algorithms have been applied to classify the emotion as a psychological measurement. However, good accuracy in measurements is yet to be achieved. This work proposes an approach to classify the different type of estimation in human being with improved accuracy. This work extracts several features from EEG brain signals and then optimize the set of features using the correlation matrix, information gain calculation, and recursive feature elimination method. The optimized feature set is then used in classification. For the classification technique, the Extreme Gradient Boosting (XGBoost) algorithm has been followed. The XGBoost algorithm works on the Gradient Boosting framework and follows linear model solver and tree learning concept. The results show that feature optimization followed by the XGBoost algorithm gives better classification accuracy. The proposed approach has been tested with DEAP data set. The proposed approach also has been compared with different classification algorithms, namely Naive Bayes, KNN, C4.5, Decision Tree, Random Forest algorithms. The results substantiate the efficacy in favor of the proposed approach.