You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The binary classification threshold threshold postprocesses probabilities as either negatives (<threshold) or positives (>=threshold). This is currently a value that one would set prior to training, and is often set to its default as 0.5. If the user wanted to change the threshold of their model upon discovery of an ideal operating point, they would edit the threshold in the config after training. However, it is currently somewhat difficult to find said operating point.
In class imbalanced settings, it is often desired to set the threshold to some value != 0.5. For example, if the user wants to have high precision at the expense of some recall of a rare positive, the user may want to set a higher threshold. We want to provide two new functionalities: (1) a visualization of threshold vs. metric, where the metrics plotted are those that are threshold-dependent (i.e. accuracy, precision, recall, f1, etc.), and (2) a lightweight ability to experiment with different thresholds from the LudwigModel.predict function.
Such functionality would enable users to analyze and ultimately select the optimal threshold for their use case.
The text was updated successfully, but these errors were encountered:
The binary classification threshold
threshold
postprocesses probabilities as either negatives (<threshold
) or positives (>=threshold
). This is currently a value that one would set prior to training, and is often set to its default as 0.5. If the user wanted to change the threshold of their model upon discovery of an ideal operating point, they would edit the threshold in the config after training. However, it is currently somewhat difficult to find said operating point.In class imbalanced settings, it is often desired to set the threshold to some value != 0.5. For example, if the user wants to have high precision at the expense of some recall of a rare positive, the user may want to set a higher threshold. We want to provide two new functionalities: (1) a visualization of
threshold
vs.metric
, where the metrics plotted are those that are threshold-dependent (i.e. accuracy, precision, recall, f1, etc.), and (2) a lightweight ability to experiment with different thresholds from theLudwigModel.predict
function.Such functionality would enable users to analyze and ultimately select the optimal threshold for their use case.
The text was updated successfully, but these errors were encountered: