The Prediction Result records [ml_predictor_results] are generated via the Business Rule Update Prediction Results, which also calls the Script Include MLPredictor, so that when the incident is closed, it creates the record with the Predicted Output Value and the Final Output Value.
ML training for Predictive Intelligence pre-processes the training data before it is trained such as removing a predefined list of stopwords, as well as special characters and spaces.
Therefore, if an Output field generates a class that contains leading or trailing spaces, these spaces will be removed before training and so the trained class will not match with the actual class on the instance, which creates the "false negative" records in the Prediction Results table, because the Predicted Output Value has the class with a name containing no leading or trailing spaces, whereas the Final Output Value contains the same class, but the name has the leading or trailing space and so generates the "false negative" Prediction Result records.
Ideally the class column values should not have any leading or trailing spaces, as we do not expect these leading or trailing spaces in names, as they can cause issues with ML training, hence we remove them in the pre-processing stage.
For example, you have created a Classification Solution for predicting the Assignment Group for a new incident. The Assignment Group value in the record is "Network ", with an extra space at the end, whereas the predicted value is "Network" without this space.
Due to this extra trailing space in the Assignment Group record, the values in Predicted Output Value and the Final Output Value do not match, hence the Predicted Correctly field is set to false. Even though the prediction was correct, the names did not match.
To fix this issue in this example, you will need to remove the trailing space for the Assignment Group and change it to "Network". Once you have removed any leading or trailing spaces in your class names, it will match the Predicted Output Value and you will no longer have "false negatives" for this class in your Prediction Results.