Interval data clustering using self-organizing maps based on adaptive Mahalanobis distances

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• High-dimensional data presents challenges in analysis and visualization.
• Artificial neural networks, like self-organizing maps (SOMs), are used for dimensionality reduction.
• Clustering interval-valued data requires specialized techniques to handle uncertainty.

Purpose of the Study:

• To develop a self-organizing map specifically for interval-valued data.
• To achieve topology preservation during the clustering of interval data.
• To introduce adaptive Mahalanobis distances for improved clustering performance.

Main Methods:

• Proposed two batch training algorithm-based methods for self-organizing maps.
• Method 1: Employed a common Mahalanobis distance for all clusters.
• Method 2: Utilized a common Mahalanobis distance per cluster, adapting to different distances during training.

Main Results:

• The proposed methods demonstrated effective clustering of interval-valued data.
• Topology preservation was maintained throughout the clustering process.
• Comparative analysis showed the advantages of adaptive Mahalanobis distances for specific datasets.

Conclusions:

• The novel self-organizing map effectively clusters interval-valued data while preserving topology.
• Adaptive Mahalanobis distances offer enhanced clustering accuracy compared to common distances.
• The proposed methods provide a valuable tool for analyzing complex interval data.