Epilepsy is a chronic brain disease characterized by recurrent seizures. These episodes are usually triggered by abnormal firing of neurons in the brain and appear as brief, recurring episodes. From the point of view of mitigation and treatment, a major problem is timely and accurate classification of epileptic seizure episodes, which is important for rapid intervention and possible prevention of future episodes. In recent years, with the rapid development of artificial intelligence technology, its application in the field of epilepsy diagnosis is increasingly extensive. In particular, electroencephalography (EEG), as a non-invasive neurophysiological examination method, plays an important role in the diagnosis of epilepsy. In particular, it allows the learning of connectomic information from the relative positions of EEG electrodes on the scalp. In this paper, a deep learning model, Temporal-Spatio-Importance Correlation Network (TSICNet), is proposed to classify the onset of epileptic seizures. Our model design combines a variety of neural network components and training methods, such as graph, spatio-temporal and separable convolutional networks, as well as sliding windows data extraction techniques and ATCNet’s time-sensitive attention mechanism, to realize effective recognition and extraction of epilepsy data features. In particular, TSICNet makes heavy use of learned connectomic information to enhance classification accuracy. Experimental results show excellent performance on the CHB-MIT Scalp EEG dataset and HUH neonatal epilpsy dataset, and has good results in accuracy and True (TPR) and False Positive Rates (FPR). Moreover, ablation analysis shows that each part of the model contributes significantly, thus confirming the validity our design principles.