Prototype Generation Using Self-Organizing Maps for Informativeness-Based Classifier.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• The k-nearest neighbor (kNN) algorithm is a fundamental data classification method.
• kNN's simplicity is offset by weaknesses including the need for exhaustive comparisons and challenges in selecting the optimal 'k' parameter in overlapping data regions.

Purpose of the Study:

• To address the limitations of the standard kNN algorithm.
• To propose a hybrid approach enhancing classification accuracy, particularly in data with poorly defined class borders.

Main Methods:

• A hybrid algorithm integrating Self-Organizing Maps (SOM) with an information-based nearest neighbor classifier (iNN).
• SOM's vector quantization is utilized for prototype generation, creating a reduced training dataset.
• The iNN classifier employs an informativeness measure for classification based on the reduced dataset.

Main Results:

• The proposed SOM-iNN combination was extensively tested.
• The hybrid approach demonstrated significant accuracy improvements on datasets with ambiguous class boundaries.
• This method effectively handles data where object classes are not well-defined in the overlap regions.

Conclusions:

• The SOM-iNN hybrid algorithm offers a robust solution to kNN's inherent weaknesses.
• This approach is particularly effective for classification tasks involving complex and overlapping data distributions.
• The method enhances classification performance by intelligently reducing the training dataset and employing an informative similarity measure.