Dual generative adversarial networks based on regression and neighbor characteristics

Area of Science:

• Machine Learning
• Pattern Recognition
• Data Science

Background:

• Imbalanced data, where sample distribution varies significantly, poses challenges for machine learning models.
• Existing research primarily addresses imbalanced classification, leaving imbalanced regression under-explored.
• Model performance is often biased towards majority samples in imbalanced datasets.

Purpose of the Study:

• To address the under-researched problem of imbalanced regression data.
• To develop novel models capable of learning both regression and neighbor characteristics.
• To improve the effectiveness of machine learning models on imbalanced regression tasks.

Main Methods:

• Introduced the Recurrent Neural Gated Unit (RNGRU) model for simultaneous learning of regression and neighbor characteristics.
• Employed adversarial learning to determine the optimal proportion of regression and neighbor characteristics.
• Proposed a correlation change rate (ccr) index to evaluate sample similarity.
• Developed the Regression-based GAN (RNGAN) model to reduce sample similarity using adversarial methods.

Main Results:

• The RNGRU model effectively learns regression and neighbor characteristics from imbalanced data.
• The RNGAN model successfully reduces the similarity between generated and original samples.
• The proposed ccr index provides a reliable measure of sample similarity.

Conclusions:

• The RNGRU and RNGAN models offer a promising approach to tackling imbalanced regression problems.
• These models enhance machine learning performance by mitigating bias towards majority samples.
• Further research in imbalanced regression is crucial for advancing pattern recognition.