Add numerical perturbation detector #2040
Conversation
|
@kevinmessiaen Please review |
|
Hey @henchaves - Would you have some bandwidth to review this PR ? |
henchaves
added
Lockfile
|
@henchaves Did you get a chance to look at this? Do I need to make any changes? |
|
Hello @Kranium2002, |
There was a problem hiding this comment.
Hi @Kranium2002, thanks so much for the contribution! I think numerical perturbation is a great addition to the library.
I've noted a few issues with the proposed implementation, in summary:
- We should check for features not only based on their data type, but column type (sometimes we have categorical features represented as numbers, but we don't want to apply the numerical perturbation to those)
- The perturbation functions used should be defined as instances of
TranformationFunction(the same we use for text perturbation or other transformations) - We need to be careful not to alter the underlying data type. Certain transformations should be applied depending on whether the feature has integer or float values (e.g. we don't want to add Gaussian noise to integers)
I also feel there is a lot of redundancy between the BaseNumericalPerturbationDetector calss and BaseTextPerturbationDetector. It's probably worth refactoring this to have a common base class covering the shared behavior instead of duplicating code — @henchaves can help with that ;)
|
|
||
| def run(self, model: BaseModel, dataset: Dataset, features: Sequence[str]) -> Sequence[Issue]: | ||
| """Run the numerical perturbation detector.""" | ||
| numerical_features = [f for f in features if pd.api.types.is_numeric_dtype(dataset.df[f])] |
There was a problem hiding this comment.
This check is not enough, because sometimes categorical features are represented with numerical types (e.g. integers). The more reliable way would be to extract this from the dataset.column_types.
| numerical_features = [f for f in features if pd.api.types.is_numeric_dtype(dataset.df[f])] | |
| numerical_features = [f for f in features if dataset.column_types[feature] == "numeric"] |
| "failed_size": failed_size, | ||
| "slice_size": slice_size, | ||
| "threshold": threshold, | ||
| "output_sensitivity": output_sensitivity, |
There was a problem hiding this comment.
add params metric, metric_value (these are used in standardized export, e.g. AVID report)
| "output_sensitivity": output_sensitivity, | |
| "output_sensitivity": output_sensitivity, | |
| "metric": "Fail rate", | |
| "metric_value": fail_rate, |
| features=[feature], | ||
| meta={ | ||
| "feature": feature, | ||
| "perturbation_fraction": self.perturbation_fraction, |
There was a problem hiding this comment.
Add domain and deviation (these are used for visualization in the scan widget).
| "perturbation_fraction": self.perturbation_fraction, | |
| "perturbation_fraction": self.perturbation_fraction, | |
| "domain": f"Feature `{feature}`", | |
| "deviation": f"{failed_size}/{slice_size} tested samples ({round(fail_rate * 100, 2)}%) changed prediction after perturbation", |
| group=self._issue_group, | ||
| level=issue_level, | ||
| description=desc, | ||
| features=[feature], |
There was a problem hiding this comment.
Add transformation_fn to indicate which transformation (i.e. the perturbation) was performed.
| features=[feature], | |
| features=[feature], | |
| transformation_fn="Numerical perturbation", # TODO: define proper transformation functions |
| lambda x: x * 1.01, | ||
| lambda x: x * 0.99, | ||
| lambda x: x + np.random.normal(0, 0.01, x.shape), |
There was a problem hiding this comment.
These should be proper transformation functions (extending TransformationFunction).
IMPORTANT: all of these perturbations assume that the numerical values are floats. This is not always the case, integer features are common: when such data types are present we must not alter that. Applying the transformations above (for example Gaussian noise) to an integer feature like num_passengers would silently convert the data type to float, potentially breaking the model inference.
Description
This PR adds the numerical perturbation detector feature to robustness detector. This PR is still a work in progress.
The detector will perturb the values by 1 percent and then check for errors in classification mode and for regression the model flags any error which is 5% away from the real value. This could be made so that the user can adjust these percentages but for now this is fixed. Review needed on this.
Related Issue
closes #1846
Type of Change
Checklist
CODE_OF_CONDUCT.mddocument.CONTRIBUTING.mdguide.pdm.lockrunningpdm update-lock(only applicable whenpyproject.tomlhas beenmodified)