This browser is no longer supported.
Upgrade to Microsoft Edge to take advantage of the latest features, security updates, and technical support.
Using Azure ML Designer it is easy to create a model using the Two-Class Logistic Regression & Train Model components. However it does not seem to be possible to view the regression coefficients / intercept (ie. the weights applied to the feature values within the model). How can we go about viewing the model coefficients? Are they stored in one of the Train Model output files (eg. data.ilearner) in a way that can be viewed / exported to a human readable format?
Note: this question relates to the Azure Machine Learning Studio (not the older 'classic' version where I believe it was possible to 'right-click' and visualise the model coefficients).
Thanks for reaching out. These are the metrics reported when evaluating binary classification models. You can view the results by clicking evaluate model > visualize > evaluation results. Hope this helps.
I'm aware of those evaluation metrics (Accuracy, F1, AUC etc) but that doesn't answer the question of how the logistic model regression coefficients can be viewed (ie. the feature weights that are added together plus the intercept to create the model's score prediction). This is a standard output of any regression model, and is required if you were to apply the model to new data outside of Azure. Many statistical tools store / output this as a PMML file (a standard XML file type that contains all the scoring rules). I hope this clarifies the question.
Any update on this? Is it still not possible to get the coefficients for the trained linear regression model?
You could use the Execute Python Script component to write code that gets the run context, which returns the run of the execute python script component, from there you get the parent of that run which is the pipeline run context. Next, iterate through the pipeline run context to find the child that contains the name "Train Model". This child has the output of the trained model, which you can download to a runtime directory, and then load the model using that runtime path. Once the model is loaded, you can access the coefficients.
Thanks for clarifying. Currently, there's no option for viewing the model coefficients in designer. However, this feature is on our roadmap and you can expect updates about this request soon. Sorry for the inconvenience.
Thanks for confirming there is currently no ability to view the model coefficients in designer. I'm glad this is on the roadmap, as this does seem like a very odd omission as it makes it hard to interpret the model and impossible for it to be 'portable' outside of Azure ML.