search

Trusted-AI / AIF360

A comprehensive set of fairness metrics for datasets and machine learning models, explanations for these metrics, and algorithms to mitigate bias in datasets and models.
https://aif360.res.ibm.com/
Apache License 2.0
2.45k stars 839 forks source link

make estimators and scorers sklearn compatible #58

Open adrinjalali opened 5 years ago

adrinjalali commented 5 years ago

The inprocessing algorithms are basically like an Estimator. Ideally, it should be possible to replace a classifier in a scikit-learn pipeline, with one from aif360. An example pipeline (taken from this example) looks like:

model = make_pipeline(StandardScaler(),
                      LogisticRegression(solver='liblinear'))
X_train = meps_orig_train.features
y_train = meps_orig_train.labels.ravel()

model_lr = model.fit(X_train, y_train,
                     **{"logisticregression__sample_weight":meps_orig_train.instance_weights})

But when we move to use a model such as PrejudiceRemover, we have to break the pipeline and have the model separate, since it doesn't follow the API requirements to be fit for a pipeline.

model = PrejudiceRemover(sensitive_attr=sens_attr, eta = 25.0)
scale = StandardScaler().fit(tr_dataset.features)

tr_dataset.features = scale.transform(tr_dataset.features)
model.fit(tr_dataset)

Once it can be fit in a pipeline, then we can use all the other mechanisms already available in sklearn, such as GridSearchCV to find best hyperparameters for the problem at hand.

That also brings are to the scorers. AIF36's scorers also don't fit into the scoring mechanism of sklearn. Once they do, we could use functions such as make_scorer to create a scoring function and feed it into the sklearn's GridSearchCV for instance.

This may not be a trivial task, and some useful things such as having multiple scoring functions recorded and reported by the grid search are still in discussion and not yet available in sklearn. Until then, we can provide an easy way for the users to combine antibias scoring functions with performance scoring ones and use them to choose their best pipeline.

Another point regarding the API conventions is that if the preprocessing modules also fit in the sklearn.Pipeline as a transformer, then we can put their selection also in a hyperparameter search and do the search much easier than having to manually run them one by one and going through them to find the best solution.

Right now transformers which would change the number of samples or change the output are not supported in sklearn (AFAIK), but that's also in discussion and this usecase may be a good push for it.

hoffmansc commented 5 years ago

Hi @adrinjalali. Apologies for poor communication about this. I’m reopening this issue because, after feedback from many other users, it sounds like stronger scikit-learn support (and probably pandas to a further degree) would be greatly appreciated. There are still a bunch of challenging technical issues that we would need to resolve which has stalled progress in this direction though. I and the rest of the team would love to have a chat with you and anyone else with ideas about this to see if we can move forward with integrating scikit-learn better with AIF360. This discussion is probably best served in a call of some sort which we can summarize afterward but this is up to you.

adrinjalali commented 5 years ago

Happy to see this going forward! I've been working on it on the fork I've made from the project, but as you say it's not an easy task, and there are a lot of things which are just done very differently here compared to scikit-learn.

However, for some of the challenges, we talked about them during the scikit-learn sprint in February in Paris, and some of the related issues are going forward.

I'd be happy to have a call, we can coordinate on your slack channel, where I'm always present.

hoffmansc commented 5 years ago

We had a good conversation about this today. Here’s a summary of what we discussed and next steps:

Challenges/workarounds:

Next steps:

Future considerations:

Feel free to add anything I've missed.

animeshsingh commented 5 years ago

@hoffmansc would it make more sense to do this in branch on sklearn? Or at-least create a major feature request there, which drives most of the conversations and points back here? With this we would be able to get more from sklearn community be able to look at what`s happening around fairness and bias, and be able to contribute ideas, code etc... ? @krvarshney

adrinjalali commented 5 years ago

@animeshsingh I'm not sure what you mean by a branch on sklearn. I don't think this would get in the main sklearn repo. For these projects, we use the scikit-learn-contrib collection or projects to give more visibility to those projects, and aif360, or a variation of it, would be a perfect match there.

There are some constraints that we need to satisfy before that. For instance, I don't think we should have a hard dependency on tensorflow, specially since it doesn't even support python3.7 yet, and by the time it does, python3.8 is probably out. Also, there seem to be some focus on python2.7 support here which is ditched already in many parts of the ecosystem (pandas and sklearn included). We're planning a release in the coming week and it supports python3.5+.

But in general, I like the idea of getting this closer to the scikit-learn-contrib realm.

animeshsingh commented 5 years ago

Thanks @adrinjalali for the response. The intention was not to move the core AIF360 there - but essentially opening up a feature request in sklearn community around having a default bias checker and mitigator, and then pointing back to the branch you or @hoffmansc has created from AIF360 here. This way the community there can be made aware of the effot

hoffmansc commented 5 years ago

I finally started some work on this in the sklearn-compat branch.

In the README for aif360.sklearn I put together a rough to-do list/roadmap. A lot of it is straightforward and I think almost all of it can be done with tricks/workarounds but explicit scikit-learn support for some of these things would be better.

Feedback and contributions are needed and we should probably advertise this in some way in case other people want to contribute.

animeshsingh commented 5 years ago

Thanks @hoffmansc. @adrinjalali it would be great to sync up once on this