A Model Based on LSTM and Data Augmentation for Network Traffic Prediction with Few Samples

Utilizing artificial intelligence technologies, particularly deep learning models, for modeling and predicting time series has demonstrated significant efficacy and potential across various fields. In the domain of network traffic, models can predict future trends by leveraging features of learned traffic patterns, thereby facilitating rational allocation of network resources as well as assessing network security and preventing cyber threats. However, deep learning models often require large amounts of high-quality time series data for training, and network traffic data, being somewhat private, is generally not easily accessible. In the real world, effective modeling is often hindered by the lack of sufficient data. In this paper, we propose a method Transfer Data Augmentation(TransDA) for augmenting few samples of network traffic data based on network traffic characteristics and the concept of transfer learning. Additionally, we have combined Long Short-Term Memory(LSTM) with the newly proposed Kolmogorov-Arnold Networks(KAN), replacing the original Multi-Layer Perceptron(MLP), to better capture the nonlinear relationships between sequential data points. Experimental results indicate that our proposed model achieves better performance on few-sample network traffic datasets compared to the standard models.