Loyalty-SMOTE: Data synthesis algorithm for effective imbalanced data classification | JoVE Visualize

Area of Science:

Machine Learning
Data Science
Artificial Intelligence

Background:

Imbalanced datasets pose significant challenges for machine learning model training, leading to suboptimal classifier performance.
Existing data-level methods for imbalanced data, such as interpolation and boundary oversampling, often overlook noise susceptibility.
There is a need for robust methods that can handle noise and improve generalization in imbalanced learning scenarios.

Purpose of the Study:

To propose a novel data-level algorithm, Loyalty-SMOTE, designed to effectively handle imbalanced datasets with noise.
To introduce the concepts of Loyalty and Attraction for identifying noisy data and generalizing to multiclass problems.
To evaluate the performance of Loyalty-SMOTE against existing methods using various metrics.

Main Methods:

Developed the Loyalty-SMOTE algorithm, incorporating a 'Loyalty' concept to identify and mitigate noisy data points.
Applied the Synthetic Minority Oversampling Technique (SMOTE) to oversample minority class boundary data after noise identification.
Introduced an 'Attraction' concept to extend the denoising technique for multiclass dataset challenges.
Utilized Support Vector Machine (SVM) as the base classifier for extensive experimental evaluation.

Main Results:

Loyalty-SMOTE demonstrated superior performance across multiple metrics on both binary and multiclass UCI datasets.
On 30 binary datasets, Loyalty-SMOTE achieved the highest F1-score in 87% of cases, AUROC in 97%, recall in 87%, and G-mean in 90%.
For 5 multiclass datasets, the algorithm yielded significant performance scores, indicating its effectiveness in complex scenarios.

Conclusions:

Loyalty-SMOTE offers a robust and effective solution for imbalanced learning problems, particularly those affected by noisy data.
The proposed 'Loyalty' and 'Attraction' concepts provide a novel framework for noise identification and handling in imbalanced datasets.
The algorithm's strong performance across diverse datasets validates its potential for improving machine learning model reliability and accuracy.