Fast and Accurate Clustering of Multiple Modality Data via Feature Matching

Area of Science:

• Computer Science
• Data Science
• Machine Learning

Background:

• Multiple modality clustering aims to group objects using cross-modality relationships for comprehensive descriptions.
• Existing methods often depend on accurate, yet costly and easily corrupted, sample-wise affinity measurements.
• Clustering big data with multiple modalities presents challenges in speed and accuracy due to heterogeneous gaps.

Purpose of the Study:

• To develop a novel, efficient, and accurate clustering approach for multi-modal data.
• To overcome the limitations of traditional methods relying on sample-wise affinity.
• To reduce computational cost while maintaining high clustering performance.

Main Methods:

• Proposes a method focusing on feature matching across modalities instead of sample-wise affinity.
• Calculates a feature matching matrix measuring feature-wise correlations.
• Decomposes the matching matrix into bases for feature spaces and uses joint reconstruction for sample assignment via alternating optimization.

Main Results:

• Significantly reduces computational cost by eliminating expensive sample-wise affinity estimation.
• Achieves high accuracy comparable to existing methods.
• Demonstrates superior speed and accuracy on both synthetic and real-world datasets.

Conclusions:

• The proposed feature matching-based clustering method offers an efficient and effective solution for multi-modal data.
• It successfully addresses the challenges of speed and accuracy in big data clustering.
• The approach provides a robust alternative to traditional affinity-based clustering techniques.