Method for Data Quality Assessment of Synthetic Industrial Data

Area of Science:

• Data Science
• Machine Learning
• Statistical Modeling

Background:

• Acquiring real-world sensor and industrial data for research can be challenging due to data access restrictions.
• Synthetic data generated through simulation is often the only alternative but may suffer from poor quality or limited quantity.
• Inadequate synthetic data can lead to unreliable algorithm testing and non-credible research outcomes.

Purpose of the Study:

• To develop and propose a mathematically grounded data-quality assessment for synthetic data.
• To ensure the suitability of synthetic data for multivariate prediction tasks that inform binary classification.
• To enable credible research and algorithm validation using high-quality synthetic datasets.

Main Methods:

• A novel data-quality assessment framework is introduced, focusing on the predictive power of independent variables.
• Assumptions for acceptable synthetic data quality are defined, with thresholds set by human assessors.
• The method facilitates the generation of a classification table for evaluating performance using standard metrics (e.g., AUC, ROC, F1 score).

Main Results:

• The proposed data-quality assessment provides a systematic approach to validating synthetic data for research.
• The framework allows for the comparison of results obtained from the proposed method with other classification techniques.
• An experimental case study using a synthetic dataset from the UCI repository demonstrated the method's applicability.

Conclusions:

• Mathematically grounded data-quality assessments are crucial for ensuring the reliability of research using synthetic data.
• The proposed method enhances the credibility of findings derived from simulated datasets in multivariate prediction and binary classification.
• This approach supports the broader use of synthetic data in scientific research where real data is inaccessible.