Conversation
jkfitzsimons
left a comment
jkfitzsimons
left a comment
There was a problem hiding this comment.
Please check out the comments. This PR does not implement a DP RandomForrestRegressor. Hopefully the comments clarify why but if you get really stuck let me know and I can help guide you! 👍
diffprivlib/models/forest.py
Outdated
| self.bounds = bounds | ||
| self.random_state = random_state | ||
|
|
||
| def _dp_mean(self, array): |
There was a problem hiding this comment.
This appears to never be called....
diffprivlib/models/forest.py
Outdated
| return private_mean | ||
|
|
||
| def fit(self, X, y): | ||
| super().fit(X, y) # Fit the RandomForestRegressor |
There was a problem hiding this comment.
Here you are fitting with the parent RandomForrestRegressor which means no differential privacy will be applied during the training (ie the main point). Please go back an look at how they created the RandomForrestClassifier via custom trees. Their trees randomly partition the domain (not using the gini or entropy coefficient like in the vanilla RandomForrestRegressor). Then once the random trees are created you can use your _dp_mean method on the leaves.
|
@jkfitzsimons please check now |
2 participants
No description provided.