Data Representativeness with Hyperdimensional Computing

Area of Science:

• Data Science
• Artificial Intelligence
• Machine Learning

Background:

• Data-driven approaches, including AI, necessitate rigorous dataset scrutiny.
• Current dataset assessment relies on subjective manual methods, limiting analysis of complex datasets.
• In-depth analysis, especially of subgroup intersectionality, is challenging with existing methods.

Purpose of the Study:

• To introduce DART (Data Representativeness), a tool for objective dataset representativeness assessment.
• To enable quantitative measurement of similarity between complex metadata distributions.
• To reduce the burden of manual dataset assessment by highlighting misaligned distributions.

Main Methods:

• Utilizing hyperdimensional computing for metadata distribution encoding.
• Applying hyperdimensional principles to assess distributional similarity.
• Developing a tool (DART) to quantitatively measure metadata similarity.

Main Results:

• DART accurately represents diverse attribute types (categorical, numeric).
• Quantitative similarity measurements identify distributions with significant misalignment.
• Demonstrated utility through four diverse case studies.

Conclusions:

• DART offers an objective, quantitative method for assessing data representativeness.
• The tool enhances the efficiency and depth of dataset evaluation for AI applications.
• DART facilitates expert focus on critical data distribution discrepancies.