ITS2Graph: Graph-based generative adversarial learning for imbalanced time series classification.

Area of Science:

• Data Mining
• Machine Learning
• Artificial Intelligence

Background:

• Time Series Classification (TSC) is crucial but challenged by class imbalance.
• Traditional methods struggle with high-order dependencies, especially with scarce minority samples.
• Improving minority class representation in imbalanced TSC is a significant challenge.

Purpose of the Study:

• To propose ITS2Graph, a graph-based generative adversarial learning framework.
• To exploit high-order associations for imbalanced time series classification.
• To enhance the representation of minority classes in imbalanced TSC.

Main Methods:

• Utilizing an auto-encoder for latent time series representation extraction.
• Constructing graphs based on pairwise similarities for node classification.
• Employing a graph generator for synthesizing minority-class features and connections.
• Training a Graph Convolutional Network (GCN) discriminator to differentiate real and generated nodes.

Main Results:

• ITS2Graph effectively addresses class imbalance in time series classification.
• The framework successfully exploits high-order associations among time series.
• Experimental results on 22 real-world datasets show superior performance compared to existing algorithms.

Conclusions:

• ITS2Graph offers a robust solution for imbalanced time series classification.
• The graph-based generative adversarial approach enhances minority class recognition.
• The method demonstrates significant improvements on diverse real-world time series data.