Seizure prediction of epileptic preictal period  through electroencephalogram (EEG) signals is important  for clinical epilepsy diagnosis. However, recent deep  learning-based methods commonly employ intra-subject  training strategy and need sufficient data, which are  laborious and time-consuming for a practical system and  pose a great challenge for seizure predicting. Besides,  multi-domain characterizations, including spatio-temporal?spectral dependencies in an epileptic brain are generally  neglected or not considered simultaneously in current  approaches, and this insufficiency commonly leads to  suboptimal seizure prediction performance. To tackle the  above issues, in this paper, we propose Contrastive  Learning for Epileptic seizure Prediction (CLEP) using a  Spatio-Temporal-Spectral Network (STS-Net). Specifically,  the CLEP learns intrinsic epileptic EEG patterns across  subjects by contrastive learning. The STS-Net extracts  multi-scale temporal and spectral representations under  different rhythms from raw EEG signals. Then, a novel triple  attention layer (TAL) is employed to construct inter?dimensional interaction among multi-domain features.  Moreover, a spatio dynamic graph convolution network  (sdGCN) is proposed to dynamically model the spatial  relationships between electrodes and aggregate spatial  information. The proposed CLEP-STS-Net achieved a  sensitivity of 96.7% and a false prediction rate of 0.072/h on  the CHB-MIT scalp EEG database. We also validate the  proposed method on clinical intracranial EEG (iEEG)  database from the Xuanwu Hospital of Capital Medical  University, and the predicting system yielded a sensitivity  of 95%, a false prediction rate of 0.087/h. The experimental  results outperform the state-of-the-art studies which  validate the efficacy of our method. Our code is available at  https://github.com/LianghuiGuo/CLEP-STS-Net. Index Terms—EEG, contrastive learning, spatio?temporal-spectral dependencies, dynamic graph  convolution, triple attention, seizure prediction