Generating locally invariant explanations for machine learning
Techniques for generating explanations for machine learning (ML) are disclosed. These techniques include identifying an ML model, an output from the ML model, and a plurality of constraints, and generating a plurality of neighborhoods relating to the ML model based on the plurality of constraints. The techniques further include generating a predictor for each of the plurality of neighborhoods using the ML model and the plurality of constraints, constructing a combined predictor based on combining each of the respective predictors for the plurality of neighborhoods, and creating one or more explanations relating to the ML model and the output from the ML model using the combined predictor.
1 . A method comprising:
identifying a machine learning (ML) model and a plurality of constraints, wherein the plurality of constraints comprises:
both: (i) a least absolute shrinkage and selection operator (LASSO) constraint and (ii) an l ∞ constraint;
wherein the l ∞ constraint provides a constraint for each predictor relating to only the neighborhood for which the predictor is generated, and
wherein the l ∞ constraint forces a matching sign relating to two generated predictors relating to two neighborhoods of the plurality of neighborhoods;
generating a plurality of neighborhoods relating to the ML model and an output from the ML model, based on the plurality of constraints;
generating a predictor for each of the plurality of neighborhoods using the ML model and the plurality of constraints, based on the output from the ML model, wherein generating the predictor comprises fitting a respective predictor to each neighborhood of the plurality of neighborhoods;
constructing a combined predictor based on combining each of the respective predictors for the plurality of neighborhoods;
creating one or more explanations relating to the ML model using the combined predictor; and
reporting the one or more explanations to explain the output from the ML model.
2 . The method of claim 1 , wherein generating the predictor for each of the plurality of neighborhoods using the input ML model and the plurality of constraints comprises:
generating a first predictor for a first neighborhood of the plurality of neighborhoods based on the plurality of constraints; and
generating a second predictor for a second neighborhood of the plurality of neighborhoods based on both an output relating to the first predictor and the plurality of constraints.
3 . The method of claim 2 , wherein constructing the combined predictor based on combining each of the respective predictors for the plurality of neighborhoods comprises:
summing a first output relating to the first predictor and a second output relating to the second predictor.
4 . The method of claim 1 , wherein the plurality of constraints comprises at least one of: (i) a least absolute shrinkage and selection operator (LASSO) constraint or (ii) an l ∞ constraint.
5 . The method of claim 1 , further comprising:
receiving an input number of neighborhoods,
wherein generating the plurality of neighborhoods comprises generating the input number of neighborhoods based on the receiving the input number of neighborhoods.
6 . The method of claim 1 , wherein generating the plurality of neighborhoods comprises using random perturbation to generate the plurality of neighborhoods.
7 . The method of claim 1 , wherein generating the plurality of neighborhoods comprises at least one of: (i) generating or (ii) selecting realistic neighborhoods.
8 . The method of claim 1 , wherein the ML model comprises a neural network, the output from the ML model comprises an inference by the neural network, and the one or more explanations relate to generating the inference using the neural network.
9 . A system, comprising:
a processor; and
a memory having instructions stored thereon which, when executed on the processor, performs operations comprising:
identifying a machine learning (ML) model, and a plurality of constraints, wherein the plurality of constraints comprises:
both: (i) a least absolute shrinkage and selection operator (LASSO) constraint and (ii) an l ∞ constraint;
wherein the l ∞ constraint provides a constraint for each predictor relating to only the neighborhood for which the predictor is generated, and
wherein the l ∞ constraint forces a matching sign relating to two generated predictors relating to two neighborhoods of the plurality of neighborhoods;
generating a plurality of neighborhoods relating to the ML model and an output from the ML model, based on the plurality of constraints;
generating a predictor for each of the plurality of neighborhoods using the ML model and the plurality of constraints, based on the output from the ML model, wherein generating the predictor comprises fitting a respective predictor to each neighborhood of the plurality of neighborhoods;
constructing a combined predictor based on combining each of the respective predictors for the plurality of neighborhoods;
creating one or more explanations relating to the ML model using the combined predictor; and
reporting the one or more explanations to explain the output from the ML model.
10 . The system of claim 9 , wherein generating the predictor for each of the plurality of neighborhoods using the input ML model and the plurality of constraints comprises:
generating a first predictor for a first neighborhood of the plurality of neighborhoods based on the plurality of constraints; and
generating a second predictor for a second neighborhood of the plurality of neighborhoods based on both an output relating to the first predictor and the plurality of constraints.
11 . The system of claim 9 , further comprising:
receiving an input number of neighborhoods,
wherein generating the plurality of neighborhoods comprises generating the input number of neighborhoods based on the receiving the input number of neighborhoods.
12 . A computer program product comprising:
a computer-readable storage medium having computer-readable program code embodied therewith, the computer-readable program code executable by one or more computer processors to perform operations comprising:
identifying a machine learning (ML) model, an output from the ML model, and a plurality of constraints, wherein the plurality of constraints comprises:
both: (i) a least absolute shrinkage and selection operator (LASSO) constraint and (ii) an l ∞ constraint,
wherein the l ∞ constraint provides a constraint for each predictor relating to only the neighborhood for which the predictor is generated, and
wherein the l ∞ constraint forces a matching sign relating to two generated predictors relating to two neighborhoods of the plurality of neighborhoods;
generating a plurality of neighborhoods relating to the ML model, based on the plurality of constraints;
generating a predictor for each of the plurality of neighborhoods using the ML model and the plurality of constraints;
constructing a combined predictor based on combining each of the respective predictors for the plurality of neighborhoods; and
creating one or more explanations relating to the ML model using the combined predictor.
13 . The computer program product of claim 12 , wherein generating the predictor for each of the plurality of neighborhoods using the input ML model and the plurality of constraints comprises:
generating a first predictor for a first neighborhood of the plurality of neighborhoods based on the plurality of constraints; and
generating a second predictor for a second neighborhood of the plurality of neighborhoods based on both an output relating to the first predictor and the plurality of constraints.
14 . The computer program product of claim 12 , further comprising:
receiving an input number of neighborhoods,
wherein generating the plurality of neighborhoods comprises generating the input number of neighborhoods based on the receiving the input number of neighborhoods.