A simple plug-in bagging ensemble based on threshold-moving for classifying binary and multiclass imbalanced data.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Class imbalance is a significant challenge in classification tasks, often addressed by rebalancing data through undersampling or oversampling.
• Existing rebalancing methods can introduce biases and require pre-specification of performance metrics.
• Threshold moving offers an alternative, allowing for adaptive performance measure tuning post-learning.

Purpose of the Study:

• To investigate the combination of bootstrap aggregating (bagging) ensembles with threshold moving for imbalanced classification.
• To demonstrate the competitiveness of this combined approach against traditional rebalancing methods.
• To extend the method for handling multiclass classification problems.

Main Methods:

• Utilized a bagging ensemble strategy without altering the natural class distribution of the training data.
• Incorporated threshold moving as a post-processing step to optimize classifier performance.
• Extended the methodology to accommodate multiclass datasets and validated using decision trees and neural networks as base classifiers.

Main Results:

• The proposed bagging with threshold moving method demonstrated competitive performance on binary and multiclass benchmark datasets.
• This approach preserves the natural class distribution, leading to well-calibrated posterior probabilities.
• The method offers flexibility in adapting to different performance measures post-hoc.

Conclusions:

• Combining bagging ensembles with threshold moving is an effective strategy for addressing class imbalance in classification.
• The method provides a robust alternative to resampling techniques, avoiding introduced biases and enabling adaptive performance tuning.
• The extension to multiclass problems further enhances its applicability in diverse machine learning scenarios.