Data Augmentation and Machine Learning algorithms for multi-class imbalanced morphometrics data of stingless bees.

Area of Science:

• Entomology
• Computer Science
• Machine Learning

Background:

• Accurate identification of stingless bee species is crucial for ecological and agricultural studies.
• Handling imbalanced datasets in species classification presents a significant challenge for machine learning models.
• Morphometric data offers a potential avenue for automated species identification.

Purpose of the Study:

• To evaluate the effectiveness of data balancing techniques, Synthetic Minority Oversampling Technique (SMOTE) and Adaptive Synthetic (ADASYN), in improving multiclass imbalanced data classification for stingless bees.
• To compare the performance of machine learning algorithms, Random Forest (RF) and Support Vector Machine (SVM), with and without these balancing techniques.
• To identify key morphometric variables for efficient and cost-effective stingless bee identification.

Main Methods:

• Applied SMOTE and ADASYN techniques in conjunction with RF and SVM algorithms.
• Utilized a six-class imbalanced dataset of stingless bee morphometrics.
• Evaluated model performance using metrics such as multi-class AUC, F1-score, G-mean, balanced accuracy, and sensitivity.
• Employed RF recursive feature elimination for variable importance assessment.

Main Results:

• Both SMOTE and ADASYN marginally improved the performance of RF and SVM classifiers.
• SVM generally outperformed RF, with SVM combined with SMOTE demonstrating superior results compared to ADASYN.
• SVM with SMOTE achieved a higher multi-class AUC (0.9918) and sensitivity (0.959) than SVM with ADASYN (AUC=0.9898, sensitivity=0.956).
• Most models correctly classified four out of six species, indicating minimal impact of imbalanced learning when classes are separable.

Conclusions:

• Data balancing techniques like SMOTE offer marginal improvements for RF and SVM in stingless bee classification.
• SVM with SMOTE is a highly effective approach for classifying stingless bee morphometric data.
• The study highlights the potential for automated machine learning applications in stingless bee identification, guiding future research towards key morphometric measurements.