Parallel Clusters: Visual Comparison of Embeddings Based on Multi-Scale Neighborhood Analysis

Understanding embedding relationships is crucial for neural network interpretability, structural analysis, and data exploration. However, visually comparing embeddings is challenging due to the difficulty of mentally aligning structures across views. In this paper, we address this challenge by constructing multiple hierarchies of data points from different perspectives to facilitate meaningful comparisons. We introduce Parallel Clusters (ParaClus), a visual analytics system that enables multi-scale exploration of embedding structures. This is achieved through cluster formations across embeddings and attributes, along with an adaptive thresholding mechanism that defines neighborhoods. By dynamically adjusting this threshold, users can explore structures at varying scales. Our interface employs a parallel axes design, where clusters derived from the same embedding or attribute are aligned along individual axes. This layout allows users to easily compare neighboring clusters and observe relationships across embeddings. Furthermore, interactive subdivision mechanisms enable users to refine clusters based on their connections to other clusters, providing deeper insights into structural dependencies. Additionally, our system seamlessly integrates labels and scalar attributes into the clustering process, offering a unified approach to analyzing multi-attribute, time-varying, and network-derived embeddings. We evaluate the effectiveness of ParaClus through expert assessments and case studies.