Improving K-Medoids Clustering with Z-Score Initialization: A Novel Approach to Reducing Iterations and Enhancing Performance

Widely recognized in data mining, the K-Medoids algorithm serves as a clustering technique for grouping data according to shared characteristics. Despite its effectiveness, the algorithm struggles with predictability and iteration stability, especially when dealing with large datasets. This study aims to address these challenges by optimizing the algorithm's performance. We introduce the use of the Z-Score method for initializing medoids, as it normalizes the data and identifies more representative central points. Experiments were conducted using the Whoscale Customer Dataset and the QSAR Fish Toxicity Dataset, both sourced from the UCI Machine Learning Repository. Utilizing Z-Score based initialization resulted in a sharp decline in the average iteration count, decreasing from 7 to 2 in the Whoscale Customer Dataset and from 6 to 2 in the QSAR Fish Toxicity Dataset. Clustering performance, assessed using the Davies-Bouldin Index (DBI), showed substantially lower DBI values, indicating superior clustering quality when compared to the conventional K-Medoids algorithm. These findings demonstrate that the Z-Score-based medoids initialization notably enhances the efficiency and reliability of the K-Medoids algorithm, making it particularly effective for large-scale datasets by improving clustering performance and reducing computational costs.