Abstract
Machine learning models often learn latent embedding representations that capture the domain semantics of their training data. These embedding representations are valuable for interpreting trained models, building new models, and analyzing new datasets. However, interpreting and using embeddings can be challenging due to their opaqueness, high dimensionality, and the large size of modern datasets. To tackle these challenges, we present WizMap, an interactive visualization tool to help researchers and practitioners easily explore large embeddings. With a novel multi-resolution embedding summarization method and a familiar map-like interaction design, WizMap enables users to navigate and interpret embedding spaces with ease. Leveraging modern web technologies such as WebGL and Web Workers, WizMap scales to millions of embedding points directly in users' web browsers and computational notebooks without the need for dedicated backend servers. WizMap is open-source and available at the following public demo link: https://poloclub.github.io/wizmap.
Introduction
In modern machine learning, high-dimensional embeddings serve as the foundational representation for everything from natural language to generative art. While these vectors capture rich semantic relationships, they are notoriously difficult for humans to interpret. As datasets scale into the millions, traditional visualization techniques often fail. Scatter plots become unreadable "hairballs" of overlapping points, while contour plots provide high-level summaries at the cost of local context. Furthermore, most existing tools require expensive backend infrastructure to process and serve massive datasets, creating a barrier for researchers working in local environments or computational notebooks.
WizMap is an open-source, web-based tool designed to bridge this gap. By combining a "Google Maps" style interaction model with a novel multi-resolution summarization technique, WizMap allows users to fluidly transition from a global overview of an embedding space down to individual data points entirely within the browser.
Methodology
WizMap’s ability to handle millions of points without a dedicated server rests on three technical pillars:
1. Multi-resolution Quadtree Aggregation
The system first projects high-dimensional embeddings into a 2D space using standard dimensionality reduction techniques like UMAP. To manage the resulting data, WizMap constructs a quadtree—a hierarchical data structure that recursively partitions the 2D space into tiles. This structure allows the system to load and render only the data necessary for the user's current zoom level and field of view.
2. t-TF-IDF (Tile-based TF-IDF)
To make the map semantically navigable, WizMap must provide meaningful labels for different regions. The authors adapted the classic Term Frequency-Inverse Document Frequency (TF-IDF) algorithm into "t-TF-IDF." By treating each tile in the quadtree as a "meta-document" containing the text associated with its embeddings, the system automatically extracts representative keywords for every region at every zoom level. As a user zooms in, labels dynamically shift from broad categories to specific sub-topics.
3. Serverless Scalability
WizMap leverages modern web technologies to move the computational burden to the client side. WebGL is used for high-performance rendering of millions of points, while Web Workers handle data processing in the background to keep the user interface responsive. This architecture makes WizMap "notebook-ready," allowing data scientists to visualize their embeddings directly within a Jupyter or VS Code environment without setting up external databases.
Case Studies and Discovery
The utility of WizMap was demonstrated through two large-scale datasets:
- ACL Anthology: Visualizing approximately 63,000 natural language processing paper abstracts from 1980 to 2022. Using WizMap, researchers could visually track the evolution of the field, observing the decline of traditional "grammar" research and the explosive rise of "large language models" and "attention mechanisms."
- DiffusionDB: Visualizing 3.6 million embeddings (1.8 million text prompts and 1.8 million generated images). WizMap successfully handled this massive scale, allowing for a side-by-side comparison of input prompts versus model outputs.
One of the tool's most powerful features is "superposition," which allows researchers to overlay two different distributions. In the DiffusionDB case study, overlaying prompts (red) and images (blue) revealed that Stable Diffusion has a higher density of "art portraits" than the prompts actually requested, indicating a specific bias in the generative model's output distribution.
Discussion and Future Work
WizMap introduces a familiar mental model—the map—to the complex world of latent spaces. By adopting intuitive pan-and-zoom interactions, the tool lowers the cognitive load required to explore millions of data points. The hierarchical summarization is also highly efficient; processing 1.8 million embeddings takes less than a minute on a standard laptop.
However, the current implementation has limitations. The quadtree approach is sensitive to initial tile boundaries, which can occasionally split a natural cluster. Additionally, while the tool supports overlaying datasets, future versions could benefit from "juxtaposition" views for side-by-side comparison of local structures. The authors also envision incorporating more advanced clustering algorithms to automatically highlight anomalies or particularly dense regions of interest.
Conclusion
WizMap provides a scalable, accessible, and performant solution for embedding exploration. By removing the need for backend infrastructure and providing dynamic, semantic labeling, it empowers researchers to better understand the high-dimensional spaces that drive modern AI.
Summary updated