Recognizing cross-subject emotions based on brain imaging data, e.g., EEG, has always been difficult due to the poor generalizability of features across subjects. Thus, systematically exploring the ability of different EEG features to identify emotional information across subjects is crucial. Prior related work has explored this question based only on one or two kinds of features, and different findings and conclusions have been presented. In this work, we aim at a more comprehensive investigation on this question with a wider range of feature types, including 18 kinds of linear and non-linear EEG features. The effectiveness of these features was examined on two publicly accessible datasets, namely, the dataset for emotion analysis using physiological signals (DEAP) and the SJTU emotion EEG dataset (SEED). We adopted the support vector machine (SVM) approach and the “leave-one-subject-out” verification strategy to evaluate recognition performance. Using automatic feature selection methods, the highest mean recognition accuracy of 59.06% (AUC = 0.605) on the DEAP dataset and of 83.33% (AUC = 0.904) on the SEED dataset were reached. Furthermore, using manually operated feature selection on the SEED dataset, we explored the importance of different EEG features in cross-subject emotion recognition from multiple perspectives, including different channels, brain regions, rhythms, and feature types. For example, we found that the Hjorth parameter of mobility in the beta rhythm achieved the best mean recognition accuracy compared to the other features. Through a pilot correlation analysis, we further examined the highly correlated features, for a better understanding of the implications hidden in those features that allow for differentiating cross-subject emotions. Various remarkable observations have been made. The results of this paper validate the possibility of exploring robust EEG features in cross-subject emotion recognition.