Utilizing Nearest-Neighbor Clustering for Addressing Imbalanced Datasets in Bioengineering.

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• Imbalanced classification is prevalent in critical domains like fault diagnosis, intrusion detection, and medical diagnosis.
• Acquiring sufficient abnormal data for training is a significant challenge in these fields.
• Existing methods often struggle with the inherent data scarcity of abnormal instances.

Purpose of the Study:

• To refine the One-Class Nearest-Neighbor (OCNN) algorithm for improved imbalanced classification.
• To introduce a novel algorithm, the Location-based Nearest-Neighbor (LBNN), for one-class problems.
• To enhance outlier identification and parameter optimization in imbalanced datasets.

Main Methods:

• Replaced the inter-quartile range mechanism in OCNN with K-means with outlier removal (KMOR) for robust outlier identification.
• Optimized algorithm parameters by treating identified outliers as non-target-class samples.
• Developed the LBNN algorithm, which clusters one-class data using KMOR and determines class membership based on farthest distance and percentile calculations for test data.

Main Results:

• The LBNN algorithm demonstrated superior performance across various metrics, including precision, recall, and G-means.
• Experiments validated the algorithm on eight standard imbalanced datasets from KEEL.
• Successful application on three real-world medical imbalanced datasets confirmed its practical effectiveness.

Conclusions:

• The refined OCNN approach, particularly the LBNN algorithm, offers a significant advancement in handling imbalanced classification problems.
• LBNN provides a robust and effective solution for scenarios where abnormal data is scarce.
• The algorithm's performance indicates its potential for widespread application in critical diagnostic and detection systems.