How Deep Learning Transforms Network Bandwidth Prediction: From RNN to CNN‑RNN Hybrids

Introduction

Network communication technology is a core component of the Internet, where each device acts as a node and exchanges data over links. It underpins services like voice/video, file transfer, streaming, data centers, peer‑to‑peer sharing, social networks, and sensor networks. With the rise of Machine Learning (ML) in fields such as computer vision and speech recognition, researchers have begun applying ML to longstanding network problems, for example, bandwidth prediction and adaptive decision‑making.

Problem 1

Accurately forecasting data traffic (or bandwidth) in a network is essential for optimization. The question posed is: if we know a node’s past bandwidth variations, how can we use deep learning to predict its future bandwidth?

Analysis and Solution

The problem can be treated as a time‑series regression task. Traditional methods include simple averages, weighted averages, and ARIMA models. Modern machine‑learning approaches, such as linear regression and support‑vector regression, also work for simple cases.

However, treating each past data point as an independent sample ignores temporal dependencies. Recurrent Neural Networks (RNNs), especially LSTM, capture long‑term dependencies and have shown strong performance in sequence modeling.

Beyond pure temporal modeling, spatial information is also valuable. By representing a city’s network as a grid, a 3‑D CNN can extract spatial features, while an RNN captures temporal dynamics. Combining CNN for spatial extraction and RNN for temporal modeling yields a hybrid architecture.

Figure 1: Using RNN for bandwidth prediction

Figure 2: Using 3D‑CNN for bandwidth prediction

Figure 3: Using CNN‑RNN for bandwidth prediction

Extension and Summary

Representing network nodes as a graph and applying Graph Neural Networks (GNN) can better capture relational information than grid‑based CNNs, offering a promising direction for bandwidth prediction.

References

[1] SUN Y, YIN X, JIANG J, et al. CS2P: Improving video bitrate selection and adaptation with data‑driven throughput prediction, Proceedings of the 2016 ACM SIGCOMM Conference. ACM, 2016: 272–285.

[2] HUANG C‑W, CHIANG C‑T, LI Q. A study of deep learning networks on mobile traffic forecasting, 2017 IEEE 28th Annual International Symposium on Personal, Indoor, and Mobile Radio Communications (PIMRC). IEEE, 2017: 1–6.

[3] WANG X, ZHOU Z, XIAO F, et al. Spatio‑temporal analysis and prediction of cellular traffic in metropolis, IEEE Transactions on Mobile Computing, 2018.