Emotion recognition from electroencephalography (EEG) signals is challenged by substantial inter-subject variability and the need for a relatively large number of channels, limiting generalizability and the development of low-channel systems. In this study, we propose an interpretable domain-adversarial framework for emotion recognition using sparse EEG channels. Following preprocessing, 111 time-domain, frequency-domain, time-frequency, and nonlinear features were extracted from each channel and reduced using the ReliefF algorithm to 43 and 68 features for the ECSMP and GAMEEMO datasets, respectively. A domain-adversarial neural network (DANN) equipped with a gradient reversal layer was then employed to reduce inter-subject variability, and model performance was evaluated using five-fold subject-level cross-validation. The proposed method achieved an accuracy of 99.40% on GAMEEMO, approximately 4% higher than the best reported result in previous studies. On the more challenging ECSMP dataset, an accuracy of 96.43% was achieved using only seven frontal EEG channels without auxiliary signals, exceeding the reported accuracy of the compared multimodal approach. SHAP analysis indicated that several wavelet-based, temporal, and amplitude-related features, particularly from frontal channels, made substantial contributions to the model's decisions. Overall, combining ReliefF-based feature selection with domain-adversarial learning using DANN reduced the number of features while improving emotion-recognition performance and mitigating the effects of inter-subject variability. SHAP analysis further provided feature-level interpretability by revealing the features contributing to model decisions.