How to calculate Shapley Values

Hi I'm trying to calculate shapley values by creating a model and specifying a query point.

I have a 100x5 array (samplealpha...) of predictor data and a corresponding 100x1 vector (imaxnum) of response variables.

This is my code so far:

responsevar=imaxnum;

% responsevar=imaxday;

% responsevar=rec;

% responsevar=dead;

array=[samplealpha;samplebeta;sampledelta;samplekappa;sampleTi;responsevar]';

% table=array2table(array);

model=fitrkernel(array(:,1:5),array(:,6));

a=shapley(model,array(:,1:5));

queryPoint=array(1,:);

b=fit(a,queryPoint);

plot(b)

This is the error message I'm getting:

Error using shapley/fit (line 400)

Unable to predict using the blackbox model for the specified 'QueryPoint' value.

Error in seirdshapley (line 88)

b=fit(a,queryPoint);

Caused by:

Error using classreg.learning.internal.numPredictorsCheck (line 29)

X data must have 5 column(s).

Please help thanks!!

Answer 1

[Star Strider]

I had no idea that shapley even existed until now, so I have no experience with it.

However, one possible problem is that ‘array’ is not a (100x5) array and is instead a (1x500) vector (after the transposition). I have no idea what the components of ‘array’ are, however the semicolon (;) operator will vertically concatenate them, not horizontally concatenate them. That requires a comma (,) delimiter.

array=[samplealpha;samplebeta;sampledelta;samplekappa;sampleTi;responsevar]';

If I guessed in error and ‘array’ actually is a (100x5) matrix, I’ll delete this Answer.

What size is ‘array’?

.

[Shaun Tan]

Hi,

'array' is a 100x5 matrix

Thanks

[Star Strider]

Then something else is wrong, and I would have to be able to have ‘array’ to work with, to see what the problem is with it. The code otherwise clearly works —

array = randn(10,6);                                % Normally-Distributed Random Variable Array To Use Default 'fitkernel' Assumptions
model=fitrkernel(array(:,1:5),array(:,6));
a=shapley(model,array(:,1:5))

a = 
  shapley with properties:

            BlackboxModel: [1×1 RegressionKernel]
               QueryPoint: []
           BlackboxFitted: []
            ShapleyValues: []
               NumSubsets: 32
                        X: [10×5 double]
    CategoricalPredictors: []
                   Method: 'interventional-kernel'
                Intercept: -0.6809

a.BlackboxModel

ans = 
  RegressionKernel
              ResponseName: 'Y'
                   Learner: 'svm'
    NumExpansionDimensions: 256
               KernelScale: 1
                    Lambda: 0.1000
             BoxConstraint: 1
                   Epsilon: 0.0558


  Properties, Methods

.