Demystifying the black-box, piece by piece.
This is the first iteration of our incremental explanation package. Currently, it includes two explanation methods: PFI and SAGE. Please look at the examples in the examples
directory. Please help us in improving our work by contributing or pointing to issues. We will update this iteration soon with further information.
ixai is intended to work with Python 3.8 and above. Installation can be done via pip
:
pip install ixai
>>> from river.metrics import Accuracy
>>> from river.ensemble import AdaptiveRandomForestClassifier
>>> from river.datasets.synth import Agrawal
>>> from ixai.explainer import IncrementalPFI
>>> stream = Agrawal(classification_function=2)
>>> feature_names = list([x_0 for x_0, _ in stream.take(1)][0].keys())
>>> model = AdaptiveRandomForestClassifier(n_models=10, max_depth=10, leaf_prediction='mc')
>>> incremental_pfi = IncrementalPFI(
... model_function=model.predict_one,
... loss_function=Accuracy(),
... feature_names=feature_names,
... smoothing_alpha=0.001,
... n_inner_samples=5
... )
>>> training_metric = Accuracy()
>>> for (n, (x, y)) in enumerate(stream, start=1)
... training_metric.update(y, model.predict_one(x)) # inference
... incremental_pfi.explain_one(x, y) # explaining
... model.learn_one(x, y) # learning
... if n % 1000 == 0:
... print(f"{n}: Accuracy: {training_metric.get():.3f}, PFI: {incremental_pfi.importance_values}")
1000: Accuracy: 0.785, PFI: {'age': 0.22, 'elevel': 0.18, 'zipcode': -0.07, 'salary': 0.04, 'commission': 0.05, 'loan': -0.06, 'car': 0.02, 'hyears': 0.03, 'hvalue': 0.03}
2000: Accuracy: 0.841, PFI: {'age': 0.26, 'elevel': 0.21, 'zipcode': -0.01, 'salary': 0.02, 'commission': 0.03, 'loan': -0.02, 'car': 0.02, 'hyears': 0.04, 'hvalue': 0.02}
3000: Accuracy: 0.921, PFI: {'age': 0.28, 'elevel': 0.24, 'zipcode': -0.00, 'salary': 0.00, 'commission': 0.01, 'loan': -0.01, 'car': 0.01, 'hyears': 0.01, 'hvalue': 0.00}
The documentation for ixai
can be found here.