Scalable Topological Data Analysis and Visualization for Evaluating Data-Driven Models in Scientific Applications.

Area of Science:

• Data Visualization
• Scientific Computing
• Machine Learning

Background:

• Large-scale scientific applications increasingly use machine learning (ML) models, presenting visualization challenges.
• Interpreting "black box" ML models and handling massive datasets (millions of samples) are key obstacles.
• Existing interpretability solutions often fail to scale beyond a few thousand samples.

Purpose of the Study:

• To present the first scalable solution for exploring and analyzing high-dimensional functions in scientific data analysis.
• To enable interactive exploration of both topological and geometric aspects of high-dimensional data.
• To provide high-level intuition for scientists working with complex datasets and ML models.

Main Methods:

• Developed a novel streaming neighborhood graph construction technique.
• Implemented corresponding topology computation methods.
• Introduced a new data aggregation scheme called topology-aware datacubes.

Main Results:

• Enabled interactive exploration of high-dimensional data, combining topological and geometric analysis.
• Demonstrated scalability beyond thousands of samples for ML model interpretation and data analysis.
• Successfully applied the method to use cases in high-energy-density physics and computational biology.

Conclusions:

• The proposed method offers a scalable approach to visualize and analyze high-dimensional scientific data and ML models.
• This technique facilitates crucial new insights in fields like high-energy-density physics and computational biology.
• Addresses the grand challenges in visualization for large-scale scientific machine learning applications.