Skater
Skater copied to clipboard
oracle
Python Library for Model Interpretation/Explanations
In the event that you have many features, you may want to reduce what you see in a feature importance plot, something like "top_n" or "min_importance" or "cumulative_variance_explained".
The docstring on the ```feature_importance``` method of ```FeatureImportance``` presents the following example: ``` >>> from skater.model import InMemoryModel >>> from skater.core.explanations import Interpretation >>> from sklearn.ensemble import RandomForestClassifier >>> rf...
When using rl=True, setup.sh is run, which updated the R version even if it is already installed. E.g. my 3.5.1 version gets updated to 3.5.2 This is kind of horrific,...
This issue needs to be broken down. The first implementation is just the basic support. Further improvement could be tacked here: https://github.com/datascienceinc/Skater/issues/259
* Improve on existing implementation * Enable support for handling multi-dimensional data Reference: https://staff.fnwi.uva.nl/m.a.migut/migut2015.pdf
Enable support for computing Shapley scores to quantify feature attribution(positive and negative influence) Reference: - Consistent Individualized Feature Attribution for Tree Ensembles (https://arxiv.org/abs/1802.03888) - A Unified Approach to Interpreting Model...
A statistical approach to trace the most influencing data points from the training set responsible for a given prediction Reference: - Understanding Black-box Predictions via Influence Functions(https://arxiv.org/pdf/1703.04730.pdf)
code: """ from skater.core.explanations import Interpretation from skater.model import InMemoryModel interpreter = Interpretation(X_test.values, feature_names=data.columns[:-1] ) model = InMemoryModel(gb.predict_proba, examples=X_train.values) plots = interpreter.feature_importance.plot_feature_importance(model, ascending = True) """ gb is gradient boosting...
During post pruning of TreeSurrogates some of the not so important nodes gets rejected if the agreed model's performance is not compromised. Will it be possible to add more insight...
A lot of the initial attempt at visualization was made using matplotlib keeping dev cost in mind. We should start thinking about improving the visualization with better libraries - bokeh,...