A Visual Analytics Approach for Categorical Joint Distribution Reconstruction from Marginal Projections

Area of Science:

• Data Science
• Information Visualization
• High-Dimensional Data Analysis

Background:

• Multivariate data are often available only as marginal distributions (projections), not full joint distributions.
• Standard high-dimensional visualization methods are inapplicable when data are fragmented across multiple tables.
• Existing methods for joint distribution reconstruction from marginals can be inaccurate, especially with limited data or categorical attributes.

Purpose of the Study:

• To develop a novel visual analytics approach for reconstructing full multivariate distributions from marginal data.
• To address limitations of existing iterative methods, particularly for categorical attributes and insufficient domain knowledge.
• To enable accurate visualization and analysis of high-dimensional data when only projections are available.

Main Methods:

• A Monte Carlo procedure is used to uniformly sample the solution space, generating a set of plausible joint distributions.
• A level-of-detail visualization system aids users in understanding patterns and uncertainties within the solution space.
• Interactive user input and constraint addition guide the iterative refinement process to narrow down solutions.

Main Results:

• The proposed approach effectively integrates anecdotal data and human expertise to iteratively refine joint distribution estimates.
• The visual analytics system facilitates comprehension of complex high-dimensional patterns and associated uncertainties.
• Interactive exploration allows users to progressively narrow the solution space, leading to a subset of high-confidence solutions.

Conclusions:

• This visual analytics framework offers a robust solution for reconstructing multivariate distributions from marginal data, outperforming traditional methods.
• The integration of human expertise with computational sampling provides a powerful tool for analyzing complex, fragmented datasets.
• The method is applicable to both numerical and categorical attributes, enhancing its versatility in data analysis.