Automated construction of classifications: conceptual clustering versus numerical taxonomy

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• Traditional numerical taxonomy relies on similarity metrics in predefined attribute spaces.
• Automated classification methods are crucial for handling large datasets.
• Existing methods often produce results lacking clear interpretation.

Purpose of the Study:

• To introduce and evaluate conceptual clustering as an automated classification method.
• To compare conceptual clustering against established numerical taxonomy techniques.
• To assess the interpretability and human-alignment of classification results.

Main Methods:

• Developed conceptual clustering, specifically conjunctive conceptual clustering.
• Implemented the method in the CLUSTER/2 program.
• Compared CLUSTER/2 against 18 numerical taxonomy methods on two datasets.

Main Results:

• Conceptual clustering produced easily interpretable classifications.
• Numerical taxonomy methods (14/18) yielded arbitrary and difficult-to-interpret results.
• Conceptual clustering results aligned well with human preferences.

Conclusions:

• Conceptual clustering provides a superior alternative to numerical taxonomy for automated classification.
• The method's strength lies in generating semantically meaningful and human-understandable classifications.
• This approach enhances the practical application of automated classification in diverse fields.