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
After training a particular model, there is typically a decent amount of threshold tuning in order to select the appropriate decision rule for the application. In general, the goal is to identify rules (either based on threshold, or other properties of the algorithm) that enable us to exclude the model from making bad decisions while still being applicable to the majority of the dataset.
For decision trees, one such possible rule is to exclude predictions which fall into leaves lacking a sufficient number of datapoints, where sufficient (>N) is specified by the user. For example, one might say that a leaf was added during training based on only 2 datapoints. During testing/scoring, we could skip samples which fall into this leaf to prevent low-confidence decision making.
Alternatives/nuance: it is possible that rather than raw number of samples, we may also want to consider impurity within the leaf, or perhaps some weighted metric weighing impurity against number of samples.
Feedback on whether this is something others have considered (I could not find any similar issues in the tracker) or whether there are clear drawbacks to this approach would be appreciated.
The text was updated successfully, but these errors were encountered:
After training a particular model, there is typically a decent amount of threshold tuning in order to select the appropriate decision rule for the application. In general, the goal is to identify rules (either based on threshold, or other properties of the algorithm) that enable us to exclude the model from making bad decisions while still being applicable to the majority of the dataset.
For decision trees, one such possible rule is to exclude predictions which fall into leaves lacking a sufficient number of datapoints, where sufficient (>N) is specified by the user. For example, one might say that a leaf was added during training based on only 2 datapoints. During testing/scoring, we could skip samples which fall into this leaf to prevent low-confidence decision making.
Alternatives/nuance: it is possible that rather than raw number of samples, we may also want to consider impurity within the leaf, or perhaps some weighted metric weighing impurity against number of samples.
Feedback on whether this is something others have considered (I could not find any similar issues in the tracker) or whether there are clear drawbacks to this approach would be appreciated.
The text was updated successfully, but these errors were encountered: