A data value metric for quantifying information content and utility.

Area of Science:

• Data Science
• Information Theory
• Machine Learning

Background:

• Data-driven innovation relies on massive, heterogeneous datasets.
• Assessing data quality and informativeness is crucial for effective decision support.
• Existing measures like VoI, QoI, and MI have limitations in quantifying utility for complex data.

Purpose of the Study:

• Introduce a novel information-theoretic measure, the Data Value Metric (DVM).
• Quantify the useful information content and utility of large, complex datasets.
• Determine if data expansion enhances or degrades information content for specific tasks.

Main Methods:

• Developed the Data Value Metric (DVM) based on a regularized model.
• DVM balances data analytical value (utility) with model complexity.
• Incorporated fidelity (model performance) and regularization (computational complexity) terms, inspired by information bottleneck.

Main Results:

• DVM effectively quantifies the useful information content (energy) of heterogeneous datasets.
• Demonstrated DVM's ability to capture the trade-off between analytical value and algorithmic complexity.
• Showcased DVM's utility in assessing the impact of data size and feature richness on information content.

Conclusions:

• DVM provides a robust method for evaluating data utility in data-driven innovation.
• The metric aids in optimizing dataset characteristics for various supervised and unsupervised learning tasks.
• DVM supports informed decisions on data acquisition and augmentation strategies.