Data labels are tags or fields that explain or give more information about a sample point, usually associated with an output. Data labels are critical in supervised machine learning as targets for optimizing predictive models using historical input and output data, making them the ground truth to train models for classification and regression problems. Optimized (or trained) models then apply the learned associations between inputs and output labels to make predictions on new data without labels. For example, a collection of images might have labels defined as “cat” or “dog,” that are then assigned to a set of input data (pixel values). When a supervised classifier is trained on this collection of labeled data, it it subsequently can make predictions of output labels (“cat” or “dog”) when presented with an unlabeled image.
Having accurate and complete output labels across a representative set of data is a fundamental requirement for any supervised learning application. High model performance is directly correlated with the amount and quality of labeled data for any enterprise AI application. Data labels in a holdout, or validation, data set can help evaluate the performance of the trained model to simulate how it would behave when presented with real-life data.
When faced with unlabeled data or unreliable data labels, data scientists resort to a variety of options, such as using simulations or empirical estimation to create labels, creating synthetic data from a limited set of labeled data, manually labeling the source data, or a combination of these steps.
The C3 AI® Suite provides native data integration, quality, cleansing, and normalization capabilities to ensure consistency and comprehensiveness of the training data. Through native data quality checks, data engineers and data scientists can have immediate visibility into availability and completeness of data labels, identify emerging gaps, and plan for strategies to improve data quality and coverage where needed.
The C3 AI Suite’s open architecture also allows integration with third-party APIs that can be leveraged to create synthetic data or augment existing labels.