A Distance Metric for Clustering Mixed Data Using Graph-Based Feature Influence Balancing Approach

Clustering is an unsupervised method for grouping data points based on their similarities. The quality of clustering outcomes typically relies on distance metrics, which measure the degree of separation between data points. Realworld datasets usually contain both numerical and categorical features, but existing mixed data distance measures typically use mathematical methods for numerical features and matching methods for categorical features. This approach is often lacking in capturing the information quality within features, thus resulting in poor clustering quality. To address this challenge, a new mixed data distance, called the Weighted Mixed Distance (WMD), is introduced to capture the influence of both types of features to improve the quality of mixed data clustering. WMD uses a graph-based unsupervised method to compute optimal weights for feature influences t oi mprove t he q uality ofm ixed data clustering. The proposed method is parameter-free. Experiments using Silhouette Index for evaluation show that the proposed WMD improves mixed data clustering quality.