<p>Cross-subject variability poses a fundamental challenge to the robustness of electroencephalogram (EEG)-based emotion recognition systems. While domain adaptation (DA) provides a promising solution, prevailing adversarial methods primarily focus on aligning marginal feature distributions. By neglecting class-specific distribution disparities across individuals, these approaches often lead to negative transfer and class confusion. To address this issue, we propose DC-MADA, a novel deep class-level multi-adversarial domain adaptation framework. Its core innovation lies in a paradigm of fine-grained, emotion category-level alignment. Specifically, each emotion class is modeled as an independent domain, with dedicated adversarial discriminators tasked to align class feature distributions across subjects. To manage unlabeled target data and mitigate the impact of noisy pseudo-labels, a soft probability weighting strategy is introduced. This strategy leverages probabilistic classifier outputs to dynamically guide the participation of target samples in the alignment of their most probable classes. The framework is integrated with a deep residual convolutional network designed to extract robust spatiotemporal EEG features. Evaluations on the SEED and SEED-IV benchmarks under a leave-one-subject-out protocol demonstrate the superiority of DC-MADA. The proposed framework achieves average accuracies of 88.87% (± 3.83%) and 75.47% (± 6.10%), respectively. These results enhanced cross-subject robustness, offering a solution for building affective brain-computer interfaces and mental health monitoring systems.</p>