Clustering Categorical Data Using Community Detection Techniques

Area of Science:

• Data Science
• Computer Science
• Machine Learning

Background:

• The k-modes algorithm offers efficient clustering for categorical data.
• Random initialization in k-modes often requires multiple trials for optimal results.
• Existing deterministic initialization methods improve clustering cost but vary in heuristic approaches.

Purpose of the Study:

• To propose a novel initialization strategy for the k-modes algorithm using community detection.
• To enhance the accuracy and efficiency of categorical data clustering.
• To provide an alternative to existing k-modes initialization techniques.

Main Methods:

• Constructing an unweighted graph from categorical data.
• Applying a fast community detection technique to identify cohesive node groups.
• Selecting the top-k detected communities as initial cluster centers (k-modes).

Main Results:

• CD-Clustering outperforms existing k-modes initialization methods.
• The proposed method shows superior performance in accuracy, precision, and recall across ten real-world datasets.
• Community detection provides a robust approach for initializing k-modes.

Conclusions:

• CD-Clustering offers a more effective and accurate initialization for k-modes algorithm.
• Community detection is a promising perspective for categorical data clustering.
• The method enhances the reliability of k-modes clustering for real-world applications.