We propose l(1) norm regularized quadratic surface support vector machine models for binary classification in supervised learning. We establish some desired theoretical properties, including the existence and uniqueness of the optimal solution, reduction to the standard SVMs over (almost) lin-early separable data sets, and detection of true sparsity pattern over (almost) quadratically separable data sets if the penalty parameter on the l(1) norm is large enough. We also demonstrate their promising practical efficiency by con-ducting various numerical experiments on both synthetic and publicly available benchmark data sets.
Electroencephalography (EEG)-based emotion recognition holds promise for real-time mental health monitoring, adaptive interfaces, and affective computing. However, accurate prediction across individuals remains challenging due to inter-subject variability and the non-stationary nature of EEG signals. To address this, we propose EMO-CARE, a lightweight deep learning framework that integrates multi-scale temporal convolutional networks with feature-level self-attention operating on multi-scale temporal representations. This architecture captures emotional patterns across diverse neural timescales while adaptively weighting multi-scale temporal features based on their relevance. Evaluated under the rigorous Leave-One-Subject-Out (LOSO) protocol on three benchmark datasets: SEED, SEED-V, and DREAMER, EMO-CARE achieves state-of-the-art accuracy with low inference latency. Extensive ablation experiments demonstrate the contribution of each architectural component, and the learned attention patterns align with known emotion-related neural activity. These findings collectively highlight EMO CARE's effectiveness in achieving subject-independent generalization and real-time applicability for EEG based emotion recognition.