Skip to content

Prediction Outcomes

Based on the predicted class and the label, Azimuth provides Outcomes, an assessment of the correctness of a given prediction. Outcomes are determined for individual utterances as a metric of success, and are also used across the application as an overall metric. These outcomes align with some business outcomes: not all errors (or correct predictions) are equal.

Rejection Class

Many ML problems have the notion of a rejection class, sometimes called the abstention class. This is the class that indicates the absence of a prediction. In some use cases, the model can predict it directly as one of its classes, and in other scenarios, it is only predicted when the confidence of a model is below a given threshold. In Azimuth, the rejection class needs to be defined in the config, as explained in the Project Configuration.

From a business outcome, it is usually less costly to not predict anything than to predict the wrong class. This is why the notion of rejection is present in the outcomes, as explained below.

Outcomes

Azimuth defines 4 different outcomes:

  • Correct & Predicted
    • When the predicted class matches the label, and is not the rejection class.
    • These are the predictions that add the most value.
  • Correct & Rejected
    • When the predicted class matches the label, but is the rejection class.
    • These predictions are correct, but do not necessarily bring a lot of value, given that the data sample does not come from one of the more meaningful classes.
  • Incorrect & Rejected
    • When the predicted class does not match the label, but the rejection class has been predicted.
    • These predicted are incorrect, but not as costly as when the wrong class is predicted.
  • Incorrect & Predicted
    • When the predicted class does not match the label, and is not the rejection class.
    • These predicted are incorrect and are the most costly.

The colors of the outcomes are used throughout the application to show how many utterances of each of the outcomes are present in different aggregations.