- Notifications
You must be signed in to change notification settings - Fork0
Post-hoc prototype-based explanations with rules for time-series classifiers
License
NotificationsYou must be signed in to change notification settings
sbobek/tsproto
Folders and files
| Name | Name | Last commit message | Last commit date | |
|---|---|---|---|---|
Repository files navigation
Post-host prototype-based explanations with rules for time-series classifiers.
Key features:
- Extracts interpretable prototype for any black-box model and creates a decision tree, where each node is constructed from the visual prototype
- Integrated with SHAP explainer, as a backbone for extraction of interpretable components (However, SHAP can be replaced with any other feature-importance method)
TSProto can be installed from eitherPyPI or directly from source code from this repository.
To install form PyPI:
pip install tsprotoTo install from source code:
git clone https://github.com/sbobek/tsprotocd tsprotopip install .For full examples on two illustrative cases go to:
- Example of extracting sine wave prototype and explaining class with existence ora absence of a prototype:Jupyter Notebook
- Example of extracting sine wave as a prototype end explaining class by difference in frequency of a prototypeJupyter Notebook
The basic usage of the TSProto assuming you have your model trained is straightforward:
fromtsproto.modelsimport*fromtsproto.utilsimport*#assuming that trainX, trainy and model are givenpe=PrototypeEncoder(clf,n_clusters=2,min_size=50,method='dtw',descriptors=['existance'],jump=1,pen=1,multiplier=2,n_jobs=-1,verbose=1)trainX,shapclass=getshap(model=model,X=trainX,y=trainy,shap_version='deep',bg_size=1000,absshap=True)#The input needs to be a 3D vector: number of samples, lenght of time-series, number of dimensions (features)trainXproto=train.reshape((trainX.shape[0],trainX.shape[1],1))shapclassXproto=shapclass.reshape((shapclass.shape[0],shapclass.shape[1],1))ohe_train,features,target_ohe,weights=pe.fit_transform(trainXproto,shapclassXproto)im=InterpretableModel()acc,prec,rec,f1,interpretable_model=im.fit_or_predict(ohe_train,features,target_ohe,intclf=None,# if intclf is given, the funciton behaves as predict,verbose=0,max_depth=2,min_samples_leaf=0.05,weights=None)
After the Interpretable model has been created it now can be visualised.
# Visualize modelfromtsproto.plotsimport*ds_final=ohe_train.copy()dot=export_decision_tree_with_embedded_histograms(decision_tree=interpretable_model,dataset=ds_final,target_name='target',feature_names=features,filename='synthetic',proto_encoder=pe,figsize=(6,3))fromIPython.displayimportSVG,ImageImage('synthetic.png')
More details on how the TSProto works and evaluation benchmarks can eb found in the following paper:
Comming soon