WISEST: Weighted Interpolation for Synthetic Enhancement Using SMOTE with Thresholds

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Imbalanced learning poses challenges in identifying rare but critical events due to classifier bias towards majority classes.
• This bias leads to the underperformance of machine learning models on minority classes, impacting real-world applications.

Purpose of the Study:

• To introduce WISEST, a novel locality-aware weighted-interpolation algorithm for generating synthetic minority samples.
• To evaluate WISEST's effectiveness in improving minority class detection on a wide range of imbalanced datasets.

Main Methods:

• WISEST employs a locality-aware weighted-interpolation approach to synthesize minority samples near class boundaries.
• The algorithm was benchmarked on over a hundred real-world imbalanced datasets, including KEEL, IoT-23, and BoT-IoT, with varying characteristics.

Main Results:

• WISEST demonstrated consistent improvements in minority detection metrics, such as recall and F1 score, on approximately half of the tested datasets.
• Relative recall increased by up to 25%, and F1 score by up to 18% compared to baseline methods.
• Trade-offs were observed in accuracy and precision, varying by dataset and classifier.

Conclusions:

• WISEST is a practical and robust method for imbalanced learning when minority data distribution allows for safe synthesis.
• The algorithm shows significant potential for improving the detection of rare events in critical applications.
• No single data sampling method uniformly excels across all imbalanced datasets, highlighting the need for tailored approaches.