Hybrid Multi-Label Classification Model for Medical Applications Based on Adaptive Synthetic Data and Ensemble...

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Multi-label categorization is growing in machine learning and computer vision.
• Existing methods like SMOTE for data balancing can introduce class overlap and noise.
• There is a need for improved techniques to handle imbalanced datasets in multi-label classification.

Purpose of the Study:

• To propose an innovative technique, Adaptive Synthetic Data-Based Multi-label Classification (ASDMLC), for imbalanced multi-label classification.
• To enhance the accuracy and robustness of multi-label classification models.
• To address the limitations of traditional data balancing methods.

Main Methods:

• Utilized Adaptive Synthetic (ADASYN) sampling to generate synthetic data for minority classes based on learning difficulty.
• Applied Min-Max normalization to standardize numerical variables.
• Employed Velocity-Equalized Particle Swarm Optimization (VPSO) for effective feature selection.
• Developed an ensemble model combining Adaptive Neuro-Fuzzy Inference System (ANFIS), Probabilistic Neural Network (PNN), and Clustering-Based Decision Tree.

Main Results:

• The proposed ASDMLC model achieved a multi-label classification accuracy of 90.88%.
• This accuracy significantly outperforms existing methods such as PCT (65.57%), HOMER (70.66%), and ML-Forest (82.29%).
• The method effectively handles class overlap and noise in imbalanced datasets.

Conclusions:

• ASDMLC offers a superior approach to multi-label classification on imbalanced datasets compared to previous techniques.
• The integration of ADASYN, VPSO, and an ensemble of ANFIS, PNN, and Decision Trees leads to improved classification performance.
• The study demonstrates the potential of adaptive synthetic data generation and advanced feature selection for robust machine learning models.