Emotion recognition plays a critical role in human-computer interaction, affective computing, and intelligent health monitoring. EEG-based affect analysis, in particular, is regarded as one of the most promising technical pathways due to its high temporal resolution and its direct reflection of underlying emotional states. However, emotion-related electroencephalogram (EEG) data typically suffer from high annotation costs, substantial inter-subject variability, and pronounced distribution shifts across devices or sessions, making it difficult for conventional deep learning models to achieve stable performance under small-sample or cross-subject conditions. To address these challenges, this study proposes ConLEAD, a contrastive-learning-driven framework for EEG-based emotion recognition, which formulates emotion classification as a relational inference task between query samples and support samples. ConLEAD consists of an embedding module and a relation module: the former employs an LSTM and a projection subnetwork to learn stable and discriminative deep EEG representations, while the latter aggregates Top-T cosine similarities between each query and its class-specific support samples to produce the final category prediction, thereby establishing a relative-labeling strategy that is adaptable to varying data scales. Benefiting from query-sample augmentation, projection-space reconstruction, and selective similarity aggregation, ConLEAD demonstrates strong robustness and generalization under small-sample, cross-subject, and cross-device scenarios. Experiments conducted on the DEAP (binary classification) and SEED (three-class classification) datasets show that ConLEAD surpasses multiple mainstream baselines (e.g., ProtoNet), achieving accuracy improvements of 8.29% / 7.35% (Valence / Arousal) on DEAP and 6.38% on SEED, validating its effectiveness and practical value for EEG-based emotion recognition.
Gao et al. (Fri,) studied this question.