First in a Series: Mastering Data Quality for Safe & Scalable AI
Saibal Banerjee, PhD and Steven Atneosen, JD
Introduction
Accuracy is the most commonly cited metric in machine learning. But if your dataset is imbalanced, accuracy can be dangerously misleading. For professionals working with sensitive PII, this misdirection can be catastrophic. Understanding how your models behave across all classes is critical for both model performance and data privacy.
Why Accuracy Fails in Imbalanced Data
A simple example can explain the importance of a balanced test set. Let’s say you have a model that labels animal photos as “dog,” “cat,” or “pig.” The green columns in Table 1 show where the actual and predicted class values agree, while the reds show where they disagree. The table below shows how to define the accuracy metric of the “What animal is this?” classifier. If 993 of 1000 images are cats, a model that always guesses “cat” yields an outstanding accuracy metric of 99.3%—yet is completely useless. Accuracy doesn’t account for class balance.
Accuracy is the simplest of all classification performance metrics. It is an easily computed and well-understood performance metric. However, it measures overall performance, independent of the classes, and is only useful when the class populations are balanced, not when they are disparately large or small to one another.
The Role of Properly Generated Training Data
Data generation tools, like tomtA.ai’s True Atomic Privacy, can create balanced test datasets while preserving feature relationships. This lets you create your models based upon precise and safe data without bias or risking exposure of personally identifiable information (“PII”).
Takeaway
When creating training data with data generation tools, always question accuracy in isolation. Use it alongside metrics that are sensitive to class distribution, and rely on balanced, privacy-safe generated training datasets. Stay tuned for our next blog in the “Mastering Data Quality for Safe & Scalable AI” series: Precision vs. Recall—Choosing the Right Metric for the Job.
Request our full technical paper with the “Book a Demo” button on our website.