H3Xplore: Points of Interests Visualization with Dynamic Keyword Clustering by Elina Oikonomaki

Visualizing high-dimensional geographic data with associated unstructured textual and quantitative attributes, such as Points of Interest (POIs), is challenging. Through a formative user survey and collaborative sessions with data scientists and engineers, we identified key challenges in the analytical and exploratory tasks of such datasets, including the detection of patterns and outliers at a global scale and the characterization of distribution and correlation across regions. To address these challenges, this tool introduces a system that integrates dynamic clustering of sentence embedding vectors with hierarchical H3 binning to reduce visual complexity and enable effective data exploration and analysis. Unstructured text tags are transformed into sentence vector embeddings using GPT-3 and BERT language models to dynamically generate meaningful clusters for effective color encodings. We employ established techniques of color weaving to generate multivariate geographic visualizations and compare the results. Additionally, to support the exploration and characterization of distributions, the interface integrates a tag-query input, a table view, a barplot view, and a level of detail (LOD) for a granular view of H3 bins.