Advancements in Fairness, Robustness, and Efficiency in Machine Learning Models

The recent developments in the research area highlight a significant shift towards enhancing fairness, robustness, and efficiency in machine learning (ML) and deep learning (DL) models, particularly in regression tasks, classification, and data augmentation techniques. A notable trend is the focus on mitigating biases and ensuring fairness in ML models, with innovative approaches like BiasGuard and FairTTTS demonstrating substantial improvements in fairness metrics without compromising accuracy. Additionally, there's a growing interest in leveraging advanced data augmentation and interpolation methods to improve model performance, especially in scenarios with limited or imbalanced datasets. Techniques such as AdaPRL and ACCon are pushing the boundaries in regression tasks by incorporating uncertainty estimation and contrastive learning, respectively, to enhance model robustness and interpretability. The exploration of procedural fairness and its relationship with distributive fairness in ML models is also gaining traction, offering new insights into achieving equitable outcomes. Furthermore, the application of multi-task learning and adversarial training methods, like CODAT, is addressing challenges related to model overfitting and class-wise robustness, ensuring more reliable and fair predictions across diverse datasets.

Noteworthy Papers

  • Data Augmentation for Deep Learning Regression Tasks by Machine Learning Models: Introduces advanced DA techniques that significantly enhance DL model performance for tabular data regression tasks.
  • BiasGuard: Guardrailing Fairness in Machine Learning Production Systems: Proposes a novel approach using TTA and CTGAN to improve fairness in deployed ML systems without retraining.
  • AdaPRL: Adaptive Pairwise Regression Learning with Uncertainty Estimation for Universal Regression Tasks: A novel framework that leverages relative differences between data points and integrates deep probabilistic models for uncertainty quantification.
  • Class Optimal Distribution Adversarial Training (CODAT): A min-max training framework that improves robust fairness in adversarial training through distributionally robust optimization.
  • FairTTTS: A Tree Test Time Simulation Method for Fairness-Aware Classification: A post-processing bias mitigation method that enhances fairness and predictive performance without the need for model retraining.

Sources

Data Augmentation for Deep Learning Regression Tasks by Machine Learning Models

Reducing Proxy Discrimination

BiasGuard: Guardrailing Fairness in Machine Learning Production Systems

Regularising NARX models with multi-task learning

Towards Fair Class-wise Robustness: Class Optimal Distribution Adversarial Training

AdaPRL: Adaptive Pairwise Regression Learning with Uncertainty Estimation for Universal Regression Tasks

Procedural Fairness and Its Relationship with Distributive Fairness in Machine Learning

ACCon: Angle-Compensated Contrastive Regularizer for Deep Regression

Kriging and Gaussian Process Interpolation for Georeferenced Data Augmentation

Mitigating Algorithmic Bias in Multiclass CNN Classifications Using Causal Modeling

FairTTTS: A Tree Test Time Simulation Method for Fairness-Aware Classification

U-Fair: Uncertainty-based Multimodal Multitask Learning for Fairer Depression Detection

Built with on top of