Model agnostic contrastive explanations for structured data
A method, system, and computer program product, including generating a contrastive explanation for a decision of a classifier trained on structured data, highlighting an important feature that justifies the decision, and determining a minimal set of new values for features that alter the decision.
1. A computer-implemented method for model agnostic contrastive explanations for interpreting a deep neural network (DNN), the model agnostic contrastive explanation method comprising:
generating a contrastive explanation for a decision of a classifier trained on structured data;
highlighting an important feature that justifies the decision; and
determining a minimal set of new values for features that alter the decision,
wherein the generating uses a zeroth-order optimization to estimate a gradient of designed loss functions for pertinent positives (PP) and pertinent negatives (PN) for the contrastive explanation.
2. The computer-implemented method of claim 1 , wherein the generating generates the contrastive explanation for any classification model that has only oracle access.
3. The computer-implemented method of claim 1 , wherein the generating is only able to query class probabilities for a desired input to generate the contrastive explanation.
4. The computer-implemented method of claim 1 , wherein, for real-valued features, the generating sets a base value as a median value of the feature.
5. The computer-implemented method of claim 1 , wherein, for categorical features, the generating sets a base value as a mode for that feature.
6. The computer-implemented method of claim 1 , wherein, for categorical features, the generating uses one of a frequency map approach (FMA) and a simplex sampling approach to handle the categorical features.
7. The computer-implemented method of claim 1 , embodied in a cloud-computing environment.
8. A computer program product for model agnostic contrastive explanations for interpreting a deep neural network (DNN), the computer program product comprising a computer-readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to perform:
generating a contrastive explanation for a decision of a classifier trained on structured data;
highlighting an important feature that justifies the decision; and
determining a minimal set of new values for features that alter the decision,
wherein the generating uses a zeroth-order optimization to estimate a gradient of designed loss functions for pertinent positives (PP) and pertinent negatives (PN) for the contrastive explanation.
9. The computer program product of claim 8 , wherein the generating generates the contrastive explanation for any classification model that has only oracle access.
10. The computer program product of claim 8 , wherein the generating is only able to query class probabilities for a desired input to generate the contrastive explanation.
11. The computer program product of claim 8 , wherein, for real-valued features, the generating sets a base value as a median value of the feature.
12. The computer program product of claim 8 , wherein, for categorical features, the generating sets a base value as a mode for that feature.
13. The computer program product of claim 8 , wherein, for categorical features, the generating uses one of a frequency map approach (FMA) and a simplex sampling approach to handle the categorical features.
14. A system for model agnostic contrastive explanations for interpreting a deep neural network (DNN), said system comprising:
a processor; and
a memory, the memory storing instructions to cause the processor to perform:
generating a contrastive explanation for a decision of a classifier trained on structured data;
highlighting an important feature that justifies the decision; and
determining a minimal set of new values for features that alter the decision,
wherein the generating uses a zeroth-order optimization to estimate a gradient of designed loss functions for pertinent positives (PP) and pertinent negatives (PN) for the contrastive explanation.
15. The system of claim 14 , wherein the generating generates the contrastive explanation for any classification model that has only oracle access.
16. The system of claim 14 , wherein the generating is only able to query class probabilities for a desired input to generate the contrastive explanation.
17. The system of claim 14 , wherein, for real-valued features, the generating sets a base value as a median value of the feature.
18. The system of claim 14 , wherein, for categorical features, the generating sets a base value as a mode for that feature.
19. The system of claim 14 , wherein, for categorical features, the generating uses one of a frequency map approach (FMA) and a simplex sampling approach to handle the categorical features.
20. The system of claim 14 , embodied in a cloud-computing environment.
21. A computer-implemented method for model agnostic contrastive explanations for interpreting a deep neural network (DNN), the model agnostic contrastive explanations method comprising:
generating a contrastive explanation for a decision of a classifier trained on structured data based only on a query class probabilities for a desired input,
wherein the generating uses a zeroth-order optimization to estimate a gradient of designed loss functions for pertinent positives (PP) and pertinent negatives (PN) for the contrastive explanation.
22. A computer-implemented method for model agnostic contrastive explanations for interpreting a deep neural network (DNN), the model agnostic contrastive explanations method comprising:
providing contrastive explanations for decisions of any black box classifier that are differentiable or not learned on tabular data,
wherein a zeroth-order optimization is utilized to estimate a gradient of designed loss functions for pertinent positives (PP) and pertinent negatives (PN) for the contrastive explanation.