Clustering and classification for dry bean feature imbalanced data

Area of Science:

• Machine Learning
• Data Science
• Computer Science

Background:

• Traditional machine learning models like Decision Trees (DT), Random Forests (RF), and Support Vector Machines (SVM) exhibit limited classification performance on imbalanced datasets.
• Imbalanced data, where one class significantly outnumbers others, poses a challenge for model training and accurate prediction.
• Existing methods often struggle to effectively handle class imbalance, leading to biased models and poor generalization.

Purpose of the Study:

• To develop and evaluate a novel hybrid algorithm for improving classification accuracy on imbalanced datasets.
• To address the limitations of traditional machine learning algorithms in handling datasets with disparate class distributions.
• To enhance key performance indicators such as precision, recall, F1-score, and Area Under Curve (AUC).

Main Methods:

• The proposed algorithm integrates Borderline-Synthetic Minority Oversampling Technique (BLSMOTE) with K-means clustering.
• BLSMOTE generates synthetic samples on the boundary of the minority class to mitigate noise and improve class representation.
• K-means clustering groups data points based on similarity, further aiding in data partitioning and model training.

Main Results:

• The combined BLSMOTE + K-means + SVM algorithm demonstrated superior classification performance compared to traditional methods on the dry bean and obesity levels datasets.
• BLSMOTE + K-means + DT successfully generated decision rules for both datasets, offering interpretable insights.
• BLSMOTE + K-means + RF effectively ranked the importance of explanatory variables, providing valuable information for feature selection.

Conclusions:

• The proposed BLSMOTE + K-means hybrid approach offers a robust solution for enhancing machine learning classification on imbalanced data.
• This method improves overall predictive accuracy and provides valuable insights through decision rules and variable importance rankings.
• The findings offer scientific evidence to support decision-making processes in fields dealing with imbalanced datasets.