A Hybrid Score to Optimize Clustering Hyperparameters for Online Search Term Data
Ariana Martino, Allison M. Rossetto · 2022 IEEE International Conference on Big Data (Big Data) · 2022
Clustering is a popular means of grouping trends in datasets. When applied to natural language data like queries to online search engines, clustering provides insights on user intent, discoverability of online data, and other topics of interest to digital marketers. However, the volume, variety, and evolving nature of the data makes manually reviewing and optimizing the hyperparameters of a clustering methodology tedious and expensive. Thus, we propose a simple optimization exercise leveraging a new metric we call the Hybrid Cluster Score (HCS), which combines three common clustering validity indices: the Silhouette Coefficient, the Calinski-Harabasz Score, and the Davies-Bouldin Index. By running multiple iterations of a clustering algorithm over a dataset with varying hyperparameters and calculating the HCS, we demonstrate that you can identify strong-performing hyperparameters that can be used to cluster data from the same source over future time periods. This method is used to optimize weekly clustering of websites’ search term data as part of a Yext search analytics product. Clusters produced via HCS-optimized hyperparameters are preferable for Yext’s product use case 79 percent of the time in human-in-the-loop testing vs. the preoptimization production state.