Inherit or Copy ClassificationLinear

I'm trying to use shapley to get the Shapley values for my logistic regression model. However, I got my model using glmfit and just have the corresponding coefficients b = [-7, 1,2,3,4,5]; while shapley only accepts objects for logistic regression made by the fitclinear function. These objects are ClassficationLinear. My idea is to convert my coefficients to this proper class type. I tried

mdl = ClassificationLinear;

mdl.Beta = b;

shapley(mdl, data);

which returns the error

Error using shapley

Blackbox model must be a classification model, regression model, or function

handle

How can I convert my logistic regression model to the proper class to be used in shapley? I'll eventually have to do this for a libsvm model too.

Answer 1

[Drew]

When you do not have a classification model or regression model that is accepted by the shapley command, you can calculate Shapley values for an arbitrary model by specifying the model prediction function with a function handle, and passing that function handle to the MATLAB shapley function. See the example "Specify Blackbox Model Using Function Handle" on the Shapley doc page: https://www.mathworks.com/help/stats/shapley.html#mw_c8688099-08ba-41d5-8a6c-0785a609b341

In this case, since you have a model built with glmfit, it is probably easiest to first rebuild the model with fitglm, and then create a function handle to the predict function of the resulting GeneralizedLinearModel object, and pass that function handle to the shapley function. The "Alternative Functionality" section of the glmfit doc page https://www.mathworks.com/help/stats/glmfit.html#mw_156845ea-98b1-4232-a762-6aa05f8f0766 indicates: "glmfit is useful when you simply need the output arguments of the function or when you want to repeat fitting a model multiple times in a loop. If you need to investigate a fitted model further, create a generalized linear regression model object GeneralizedLinearModel by using fitglm or stepwiseglm. A GeneralizedLinearModel object provides more features than glmfit."

% Load data and build GLM. This model is like in the example at 
% openExample('stats/FitaLogisticRegressionModelExample')
load hospital
fulldata = dataset2table(hospital);
% Only keep the variables being used
data=fulldata(:,[2 3 4 5]);
modelspec = 'Smoker ~ Age*Weight*Sex - Age:Weight:Sex';
mdl = fitglm(data,modelspec,'Distribution','binomial');

% Create a function handle to the predict function for this model
f = @(tbl) predict(mdl,tbl);

% Calculate shapley values using the function handle.
% This can be done all in one step (commented out)
% explainer=shapley(f,data(:,[1:3]),'CategoricalPredictors',[1],'q',data(1,:));

% Or, use separate calls to 'shapley' and 'fit'.
explainer=shapley(f,data(:,[1:3]),'CategoricalPredictors',[1]);
explainer=fit(explainer,data(1,:));

% View the ShapleyValues 
explainer.ShapleyValues

ans = 3×2 table
    Predictor    ShapleyValue
    _________    ____________

    "Sex"          0.043428  
    "Age"         -0.001237  
    "Weight"       0.042887  

% Plot the Shapley Values using the explainer plot method
plot(explainer)

If this answer helps you, please remember to accept the answer.