An Integrated Cluster Detection, Optimization, and Interpretation Approach for Financial Data.

Area of Science:

• Data Science
• Financial Analytics
• Machine Learning

Background:

• Automated cluster detection is crucial for financial applications like fraud detection and credit evaluation.
• Financial data complexity often hinders accurate clustering and interpretation of user behaviors and risks.
• Existing methods struggle with complex data distributions and providing easily interpretable clusters.

Purpose of the Study:

• To develop an integrated approach for detecting and optimizing clusters in financial data.
• To enhance the interpretability of detected clusters for better risk assessment and user behavior analysis.
• To create a computationally efficient method for analyzing large-scale financial datasets.

Main Methods:

• Proposed a novel cluster quality evaluation criterion for adaptive hyperellipsoidal cluster detection.
• Developed a new solver for a revised support vector data description model to refine cluster centroids and scopes.
• Utilized k-Means as a base clustering algorithm guided by the new evaluation criterion.

Main Results:

• The proposed algorithm efficiently identifies a reasonable number of clusters in financial datasets.
• Refined clusters exhibit tighter data distributions, leading to greater similarity within clusters.
• Demonstrated effectiveness across ten diverse financial datasets, confirming efficiency and accuracy.
• Enabled easier interpretation of clusters using eigenvectors.

Conclusions:

• The integrated approach successfully detects and optimizes clusters in financial data, improving interpretability.
• The method is suitable for large-scale financial datasets with meaningful features.
• Applicable to various financial mining tasks, including data distribution interpretation and anomaly detection.