Emotion recognition is gaining increasing attention from researchers and health professionals, and it has become a hot research topic in recent years. However, there have been no fruitful achievements on emotion recognition owing to the low quality of features extracted from the original electroencephalogram (EEG) and low emotion recognition rate. In this study, we extract time–frequency domain and chaotic features from human electroencephalogram signals using wavelet packet transform and chaos theory. The two types of features are combined to generate a mixed feature set with low dimensions. Then, we propose a one-to-one long short support vector machine (LS-SVM) classifier based on the Gauss function for multiclass classification problems. Focusing on the indivisible classified regions in the decision model, we combine the distribution of samples with distance between samples and separating hyperplanes, and then establish a discriminant function for fuzzy classification in the one-to-one LS-SVM. Three groups of comparative experiments, namely, feature selection and dimensionality reduction, the effectiveness of optimized LS-SVM, and the classification of different types of feature set, are conducted. The experimental results for the mixed feature set demonstrate that the proposed classification model has a competitive performance.