Distance Metric Based Oversampling Method for Bioinformatics and Performance Evaluation.

Area of Science:

• Computer Science
• Data Science
• Machine Learning

Background:

• Imbalanced datasets present a common challenge in classification, leading to poor performance on minority classes.
• Standard classification methods often exhibit high accuracy for majority classes but fail to adequately predict minority classes.

Purpose of the Study:

• To develop and evaluate an efficient algorithm for handling imbalanced datasets.
• To improve the classification accuracy of minority classes in imbalanced datasets.

Main Methods:

• A novel algorithm coupling Top-N Reverse k-Nearest Neighbor (TRkNN) with the Synthetic Minority Oversampling TEchnique (SMOTE) was designed and implemented.
• The proposed algorithm was tested using various classification methods including logistic regression (LR), C4.5, Support Vector Machine (SVM), and Back Propagation Neural Network (BPNN).
• Different distance metrics (Euclidean, Manhattan, Chebyshev, Cosine) were employed to assess their impact on classification performance.

Main Results:

• The TRkNN-SMOTE algorithm demonstrated significant improvements in classifying minority classes across various datasets.
• Euclidean and Manhattan distances proved to be more accurate and computationally efficient than Chebyshev and Cosine distances for this task.
• The proposed algorithm effectively addressed the class imbalance problem, enhancing overall classification performance.

Conclusions:

• The TRkNN-SMOTE algorithm is a robust and efficient method for addressing imbalanced classification problems.
• The choice of distance metric significantly impacts performance, with Euclidean and Manhattan distances recommended for imbalanced datasets.
• This research provides valuable insights and recommendations for handling imbalanced data in future studies.