dylan-slack / modeling-uncertainty-local-explainability Goto Github PK

Hi Dylan!

I was running your code on a number of datapoints to calculate SHAP estimates using KernelSHAP and got "negative dimensions are not allowed" numpy error on a particular datapoint after which the code crashed. The error was raised at line 521 in explanations.py. I believe the error is because the value for n_needed for the particular datapoint came out to be lower than the value of ptg_inital_points which is set it to its default value of 200. This is not happening for all datapoints because the n_needed value caluclated each time is different of course, but adding an if-else statement to handle this edge case and bypass the second call to _shap_tabular_perturb_n_samples fixed it.

Regards
Suchismita

Hi Dylan!

The value for z for the standard normal distribution used in the calculation of the PTG estimate has been hardcoded to 1.96 on line 105 in method get_ptg in regression.py. This would give an erroneous PTG estimate for the desired confidence levels other than 95%, hence this value needs to be dynamically calculated using self.percent.

Regards
Suchismita

Hi Dylan,

When I run BayesSHAP on my regression task, I encounter the convergence error at the self.enumerate_initial_shap function. Any idea why this happens?
I saw enumerate_initial is set to True as default which leads to this self.enumerate_initial_shap. Is it necessary to do so when performing BayesSHAP?
Also is BayesSHAP scalable to high input dimensions?

Best
Robin

Traceback (most recent call last):
  File "/home/miniconda3/envs/hpmodel/lib/python3.7/site-packages/IPython/core/interactiveshell.py", line 3441, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "<ipython-input-15-414d493a31bd>", line 14, in <module>
    e_i = exp.explain(X_bayesshap[i], **exp_kwargs)
  File "/home/Robin/BNN_Torch/bayesshap/explanations.py", line 651, in explain
    l2=l2)
  File "/home/Robin/BNN_Torch/bayesshap/explanations.py", line 457, in _explain_bayes_shap
    data_init, inverse_init = self._enumerate_initial_shap(data, max_coefs)
  File "/home/Robin/BNN_Torch/bayesshap/explanations.py", line 417, in _enumerate_initial_shap
    inverse = self.shap_info.discretizer.undiscretize(data)
  File "/home/miniconda3/envs/hpmodel/lib/python3.7/site-packages/lime/discretize.py", line 145, in undiscretize
    feature, ret[:, feature].astype(int)
  File "/home/miniconda3/envs/hpmodel/lib/python3.7/site-packages/lime/discretize.py", line 132, in get_undiscretize_values
    random_state=self.random_state
  File "/home/miniconda3/envs/hpmodel/lib/python3.7/site-packages/scipy/stats/_distn_infrastructure.py", line 980, in rvs
    vals = self._rvs(*args)
  File "/home/miniconda3/envs/hpmodel/lib/python3.7/site-packages/scipy/stats/_distn_infrastructure.py", line 913, in _rvs
    Y = self._ppf(U, *args)
  File "/home/miniconda3/envs/hpmodel/lib/python3.7/site-packages/scipy/stats/_continuous_distns.py", line 7163, in _ppf
    return _truncnorm_ppf(q, a, b)
  File "/home/miniconda3/envs/hpmodel/lib/python3.7/site-packages/scipy/stats/_continuous_distns.py", line 6933, in vf_wrapper
    return vf(*args)
  File "/home/miniconda3/envs/hpmodel/lib/python3.7/site-packages/numpy/lib/function_base.py", line 2163, in __call__
    return self._vectorize_call(func=func, args=vargs)
  File "/homeminiconda3/envs/hpmodel/lib/python3.7/site-packages/numpy/lib/function_base.py", line 2246, in _vectorize_call
    outputs = ufunc(*inputs)
  File "/home/miniconda3/envs/hpmodel/lib/python3.7/site-packages/scipy/stats/_continuous_distns.py", line 7113, in _truncnorm_ppf
    maxiter=TRUNCNORM_MAX_BRENT_ITERS)
  File "/home/miniconda3/envs/hpmodel/lib/python3.7/site-packages/scipy/optimize/zeros.py", line 780, in brentq
    r = _zeros._brentq(f, a, b, xtol, rtol, maxiter, args, full_output, disp)
RuntimeError: Failed to converge after 40 iterations.

Hi Dylan,

Thanks a lot for releasing the code! I'm looking into applying your BayesSHAP to explain the input feature importance of a high-dimensional regression model. Based on my understanding on your method, the Bayesian inference is directly applicable to regression task. I've made the following minor edits to explanations.py in order to perform BayesSHAP on regression tasks.

• line 37: the input argument categorical_features is set to None as there is no categorical inputs in my case
• line 59: assert mode in ["classification"] -> assert mode in ["classification", "regression"]
• line 553: the input argument label is set to 0 as there is only 1-d output for regression
• use a regression prediction model for all classifier_f with output dimension [n_test, 1]

I'd deeply appreciate if you could double-check whether those changes are correct/sufficient. Thanks a lot.

Robin

Hello, will regression versions of Bayes LIME and SHAP be supported in the future? Thank you.