This research paper describes an original method of affective state recognition in the e-learning field. Emotion plays a key role in both knowledge-building and mental health. In recent years, more and more emotion researchers have broken through traditional questionnaires to utilize physiological data to monitor students' cognitive and affective states. Among these methods, electroencephalography (EEG) can directly reflect the physiological activities of the brain and has unique potentials and advantages over others. For analyzing this multi-channel noised signal, current emotion recognition methods mainly use deep learning methods to learn the spatial or temporal representation of each channel, and then process the classification through a multimodal fusion strategy, while emotional expression highly relies on brain functional connectivity. In this research work, for the EEG-based learning-centered affective state recognition, we adopted a novel residual shrinkage block (RSB) to construct the graph neural network (GNN). During the feature extraction, the RSB is designed to obtain the features of interest and reduce the influence of artifact noises for recognition. GNN considers the biological topology among different brain regions to capture relations among different EEG channels. Extensive experiments on the CAL dataset prove that the performance of the proposed model is superior to current deep learning methods. Prior research may use the findings of this study to empower adaptive self-regulated learning environments through the automated recommendation of learning strategies, learning contents, and emotion regulation strategies according to students' learning-centered affective states, to further improve their learning performance as well as mental health. On the other hand, teachers or online course designers can use emotional feedback to adjust the learning materials and the pace of the instruction according to students' needs and preferences.