Adapting the Right Measures for Pattern Discovery: A Unified View.

Area of Science:

• Data Mining
• Pattern Recognition
• Machine Learning

Background:

• Interestingness measures are crucial for effective pattern discovery in large datasets.
• Existing measures often lack theoretical properties suitable for efficient pattern mining.

Purpose of the Study:

• To establish necessary conditions and desirable properties for interestingness measures.
• To introduce and validate a new property, SA-CAMP, for pattern discovery.
• To develop and evaluate an algorithm, GMiner, based on the proposed property.

Main Methods:

• Formal definition of three necessary conditions for interestingness measures.
• Introduction and theoretical proof of the support-ascending conditional antimonotone property (SA-CAMP).
• Development of the GMiner algorithm leveraging SA-CAMP and generalized mean (GM) measures.
• Experimental validation of GMiner's efficiency and effectiveness.

Main Results:

• Established three necessary conditions for interestingness measures in association pattern discovery.
• Identified SA-CAMP as a key desirable property for efficient pattern mining.
• Demonstrated that measures with SA-CAMP are suitable for specific itemset-traversal structures.
• The GMiner algorithm effectively identifies interesting patterns using SA-CAMP, even at very low support levels.

Conclusions:

• SA-CAMP provides a robust theoretical foundation for developing efficient pattern discovery algorithms.
• The GMiner algorithm offers a practical and efficient solution for finding interesting association patterns.
• The proposed framework unifies the view of interestingness measures, advancing the field of pattern discovery.