Emotion recognition is an important field of research in Affective Computing (AC), and the EEG signal is one of useful signals in detecting and evaluating emotion. With the development of the deep learning, the neural network is widely used in constructing the EEG-based emotion recognition model. In this paper, we propose an effective similarity learning network, on the basis of a bidirectional long short term memory (BLSTM) network. The pairwise constrain loss will help to learn a more discriminative embedding feature space, combined with the traditional supervised classification loss function. The experiment result demonstrates that the pairwise constrain loss can significantly improve the emotion classification performance. In addition, our method outperforms the state-of-the-art emotion classification approaches in the benchmark EEG emotion dataset-SEED dataset, which get a mean accuracy of 94.62%.