Deep Generative Mixture Model for Robust Imbalance Classification.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Imbalanced data classification presents challenges for minority classes, leading to unstable predictions and poor performance in existing methods.
• Discovering hidden patterns in imbalanced datasets is crucial for real-world applications.

Purpose of the Study:

• To propose a deep generative classifier that addresses limitations in imbalanced data classification.
• To enhance prediction stability and performance through model and data perturbation.

Main Methods:

• A deep generative classifier derived from a deep latent variable model with two variables is proposed.
• One variable captures data information as probability distributions (latent codes) for model perturbation and stable predictions.
• The second variable acts as a prior, restricting latent codes to Gaussian Mixture Model components for data perturbation.

Main Results:

• Extensive experiments were conducted on widely-used real imbalanced image datasets.
• The proposed model demonstrated superior performance compared to popular imbalanced classification baselines.
• The model achieved improved results on the imbalanced classification task.

Conclusions:

• The proposed deep generative classifier effectively mitigates issues associated with imbalanced data.
• The integration of model and data perturbation leads to more stable and accurate predictions.
• This approach offers a promising solution for challenging imbalanced classification problems.