Coupled attribute similarity learning on categorical data.

Area of Science:

• Data Science
• Machine Learning
• Statistics

Background:

• Traditional categorical data similarity analysis often assumes attribute independence, which is unrealistic in real-world datasets.
• Existing methods for attribute dependency aggregation offer only a local view and suffer from high computational complexity with increasing data scale.

Purpose of the Study:

• To propose an efficient, data-driven similarity learning approach for categorical data that captures global attribute interactions.
• To develop a coupled attribute similarity measure that addresses attribute couplings and enhances unsupervised learning.

Main Methods:

• Developed a coupled attribute similarity measure integrating intra-coupled (within-attribute) and inter-coupled (between-attribute) similarities.
• Designed four inter-coupled similarity measures based on set theory (power set, universal set, joint set, intersection set).
• Proposed two new coupled categorical clustering algorithms: CROCK and CLIMBO.

Main Results:

• The intersection set-based measure demonstrated superior efficiency and equivalent accuracy, particularly for large-scale datasets.
• Experiments on 13 UCI datasets showed significant performance improvements in clustering algorithms using the coupled dissimilarity metric.
• The proposed coupled attribute similarity is effective and efficient for capturing intrinsic and global attribute interactions in large datasets.

Conclusions:

• The novel coupled attribute similarity measure effectively addresses attribute dependencies in categorical data.
• The proposed approach and algorithms offer significant improvements in clustering quality and efficiency for large-scale categorical data.
• The method is generic and applicable to various data structures and clustering tasks.