EVCLUS: evidential clustering of proximity data

Area of Science:

• Data Science
• Artificial Intelligence
• Belief Functions

Background:

• Relational clustering methods analyze object dissimilarities.
• Existing techniques like hard, fuzzy, and possibilistic clustering have limitations in representing data uncertainty.
• Dempster-Shafer theory provides a framework for reasoning under uncertainty.

Purpose of the Study:

• To introduce a new relational clustering method based on Dempster-Shafer theory.
• To develop a method that assigns belief functions to objects reflecting their dissimilarity.
• To introduce the concept of credal partitions for enhanced data analysis.

Main Methods:

• The proposed method, evidential clustering (EVCLUS), utilizes Dempster-Shafer theory.
• It assigns a basic belief assignment (mass function) to each object.
• The conflict between mass functions quantifies object dissimilarity, forming credal partitions.

Main Results:

• EVCLUS effectively assigns belief assignments reflecting object dissimilarity.
• The introduced credal partition concept offers a generalized view of clustering.
• Experiments show EVCLUS performs well compared to state-of-the-art relational clustering methods.

Conclusions:

• Evidential clustering provides a robust framework for relational data analysis.
• The method enhances understanding of data structure through credal partitions.
• EVCLUS demonstrates superior performance in relational clustering tasks.