Abstract: To address the issue of low accuracy in predicting network traffic for FlexE clients, an attention mechanism that can better capture local contextual information of network traffic sequences is designed. The node embedding method was introduced to extract personalized features of different FlexE clients. To reduce the significant increase in parameters caused by node embedding, a matrix decomposition method was proposed. We allocated FlexE calendar slots based on the prediction results and compared the performance of different resource allocation methods. Experimental results demonstrate the effectiveness of the proposed method.