Abstract: In recent years, with the rapid development of signal processing and machine learning technology, EEG-based emotion recognition has received more and more attention, in which feature extraction is a key step. This paper proposes an improved local discriminant bases (ILDB)algorithm, in which both the energy and the mean of each signal subspace coefficients are extracted from ILDB to construct feature vectors and SVM is utilized to classify. By assessing the separability of feature vectors and classification accuracy rate, the extracted features via ILDB are separable and have higher classification accuracy. The highest average classification accuracy rate of ILDB algorithm can attain 88%, which is 4. 4% and 7. 2% higher than that of LDB algorithm. Moreover, the average classification accuracy rate in all channels of ILDB algorithm increases by 10. 1% and 9. 8%.