A new possibilistic-based clustering method for probability density functions and its application to detecting...

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• Unsupervised learning excels at abnormality detection, but existing methods primarily handle discrete data, neglecting probability density functions (PDFs).
• Clustering PDFs is crucial for analyzing complex datasets, yet current techniques face limitations in identifying abnormal elements within these functions.

Purpose of the Study:

• To develop a novel possibilistic clustering algorithm for identifying abnormal elements within probability density functions.
• To enable effective abnormality detection in big data by analyzing underlying probability density functions.

Main Methods:

• Data extraction via density functions followed by a proposed possibilistic clustering algorithm.
• Generation of a possibilistic partition and establishment of a decision rule for abnormality recognition.
• Comparison with baseline clustering algorithms like k-means, FCF, and Self-Updated Clustering for PDFs.

Main Results:

• The proposed algorithm achieved 100% accuracy on simulated benchmark data, surpassing baseline methods.
• Application to image data yielded G-mean scores from 96% to 100%, with Sensitivity (92-100%) and Specificity (100%).
• Demonstrated superior performance in clustering PDFs and detecting abnormal elements compared to existing algorithms.

Conclusions:

• The developed possibilistic approach effectively clusters probability density functions and detects abnormalities.
• This method offers a robust tool for understanding internal structures within big data through PDFs.
• The algorithm shows significant potential for research and application in the digital age for advanced data analysis.