Deep learning based adaptive sequential data augmentation technique for the optical network traffic synthesis.

Area of Science:

• Optical Networks
• Machine Learning
• Data Science

Background:

• Machine learning in optical networks is hindered by insufficient and non-diverse training data.
• Effective data augmentation is crucial for improving ML model performance in optical network applications.

Purpose of the Study:

• To propose a deep learning-based sequential data augmentation technique for aggregate traffic data in optical networks.
• To evaluate the effectiveness of a generative adversarial network (GAN) model for traffic data augmentation compared to traditional methods.

Main Methods:

• A generative adversarial network (GAN) model was trained on experimental optical network traffic data.
• The GAN adaptively augmented traffic data by extracting key characteristics through zero-sum game theory.
• Comparisons were made with Statistical Parameter Configuration (SPC) and Variational Autoencoder (VAE) models.

Main Results:

• GAN-augmented data closely matched experimental data in mean and variance (within 2% deviation).
• The Hurst exponent of GAN-augmented data achieved 90-96% similarity to experimental data across network types.
• K-mean clustering demonstrated over 95% accuracy for typical traffic types using GAN-generated data.

Conclusions:

• The proposed GAN-based data augmentation effectively generates diverse and realistic optical network traffic data.
• This method requires minimal initial data and can generate unlimited augmented data.
• The technique shows potential for broader applications in sequential data augmentation for optical networks.