Trafficnet: A Novel Network Performance Prediction Model via Aggregator-Based Enhancement

Network performance estimation is a key enabler in achieving efficient network operation. It can assist in important tasks such as topology design, parameter tuning, and capacity planning. The most popular methods have recently been based on Convolutional Neural Network(CNN) or Recurrent Neural Network(RNN). However, many of these methods focus excessively on particular aspects of network features. They often overlook the diversity and complexity in network performance evaluation. This paper proposes a novel model TrafficNet for network performance estimation. This model uses aggregators to learn inter-flow correlations and intra-flow dependencies. By creating specialized feature extraction components for different types of network traffic and using the adaptive mechanism to fuse these features, we aim to improve the accuracy of network performance evaluation. Furthermore, our extensive experiments have shown that TrafficNet can improve the Mean Squared Error(MSE) by 58.3% compared with the SOTA models.