Fixing imbalanced binary classification: An asymmetric Bayesian learning approach.

Area of Science:

• Statistics
• Machine Learning
• Computational Statistics

Background:

• Standard binary classification models assume balanced data, which can lead to poor performance and bias with imbalanced datasets.
• Traditional symmetric link functions (e.g., logit, probit) may not effectively handle classification tasks with skewed data distributions.

Purpose of the Study:

• Introduce novel asymmetric link functions for binary regression based on the Lomax distribution and its variants.
• Evaluate the performance of these new functions in addressing data imbalance challenges in classification.

Main Methods:

• Developed new Bayesian asymmetric classification functions using Lomax distribution variations (power, reverse).
• Implemented these functions within an R workflow using Stan for Bayesian inference.
• Compared the proposed models against classical symmetric link functions using real-world imbalanced datasets.

Main Results:

• The proposed asymmetric Lomax functions demonstrated superior performance over traditional link functions in binary classification tasks.
• Specifically, the reverse power double Lomax (RPDLomax) model effectively distinguished between failure and success probabilities in imbalanced data.
• RPDLomomax assigned lower probabilities to failure (21.4%) and higher to success (63.7%), unlike logistic regression (36.0% failure, 39.5% success).

Conclusions:

• The proposed asymmetric Lomax approach offers a competitive and effective alternative for binary data classification, particularly in scenarios with imbalanced datasets.
• These novel functions provide better differentiation between classes compared to standard logistic regression.
• The Bayesian implementation in R facilitates practical application and further research in asymmetric binary regression.