EEG-based emotion recognition has achieved significant progress in recent years, while cross-dataset modeling remains a formidable challenge. Most existing studies focus on addressing device differences and data distribution discrepancies, but they do not sufficiently consider the issue of label inconsistency across datasets. To address this issue, this paper proposes a novel cross-dataset framework named Cognitive-Inspired Hierarchical Learning (CIHL). Inspired by the coarse-to-fine characteristics of human cognition, the framework maps tasks from different datasets into a unified coarse-grained label space to jointly learn global emotional representations, while progressively optimizing fine-grained representations for the target dataset through a hierarchical structure. Specifically, CIHL includes two key designs: (1) a progressive attention module (PAM), which models coarse-grained emotions through self-attention to capture global emotional patterns and further utilizes shared key–value representations to guide the learning of fine-grained emotions; and (2) a hierarchical label smoothing (HLS) strategy, considering that fine-grained categories within the same coarse-grained emotion are semantically closer and assigns smoothing weights to related categories during fine-grained feature optimization, thereby promoting emotion-related representation learning. Extensive experiments on the SEED-V and SEED-VII datasets demonstrate that CIHL consistently outperforms the current state-of-the-art (SOTA) methods, showing strong generalization ability and stable cross-dataset performance. Specifically, CIHL surpasses SOTA methods by 1.60% and 2.16% in average accuracy on SEED-V and SEED-VII, respectively.
Han et al. (Thu,) studied this question.