Regression Learner App - relative weights of variables

Hello everyone,

I've been using the Regression Learner App to train a model. Can anyone tell me how can I check the relative weights that the model assigns to every predictor variable?

thanks a lot,

best,

Laura

Answer 1

[Kevin Holly]

Assuming that you exported your trained model as the variable trainedModel and that you have a linear model, you can access a table that has the coefficient estimates of the predictor variables as such:

trainedModel.LinearModel.Coefficients

You could extract those values by typing:

trainedModel.LinearModel.Coefficients.Estimate

You could also determine which of the predictors seemed to have the most impact by using something like LIME (Local Interpretable Model-Agnostic Explanations).

r=lime(trainedModel.predictFcn,train_data,'type','Regression');
qp=train_data(1,:);
r2=fit(r,qp,3);
plot(r2);

[lauzof]

Hi Kevin,

thanks for your answer. It's the same procedure in the case of a Optimizable GPR (Gaussian Process Regression)?

I'm running the code you posted for the attached model (model_24Jan_443_490_560_665_resources) and training dataset (tbl.training.mat). You can see that the required variables of my model (trainedModel.RequiredVariables) are log_rrs443, log_rrs490, log_rrs560 and log_rrs665. However, in the LIME plot I see other variables not used as predictors of my model (lat, log_MBR_OC3 and log_rrs490). Any thoughts on that?

Thanks again in advance!

Laura

[Kevin Holly]

Try this:

load('TrainedRegressionModel.mat')
load('tbl_training.mat')

r = lime(trainedModel.RegressionGP,tbl_training,'Type','regression');
qp=tbl_training(1,:); % This is the query point.
r2=fit(r,qp,4); % You had 4 predictors, so I changed 3 to 4
plot(r2);

r = lime(trainedModel.RegressionGP,tbl_training,'Type','regression');
qp=tbl_training(10,:); % This is the query point.
r2=fit(r,qp,4); % You had 4 predictors, so I changed 3 to 4
plot(r2);

[lauzof]

Thanks a lot, Kevin!

Answer 2

[lauzof]

Is it possible to find the predictor importance from my model at a global level? something as shown in the first plot of here https://it.mathworks.com/help/stats/lime.plot.html but not for an individual queryPoint

thanks again!

best,

Laura