Visual machine learning using SAS Viya: a Graduate Intern’s perspective

A sample of the variety of hyperparameter modifications available.

SAS VDMML creates a simplified approach to machine learning solutions beneficial to people with a wide range of expertise.

Have you been programming for as long as you can remember and are well-versed in the machine learning world?

Why spend all your precious mental energy focusing on tedious programming tasks? Instead, you could be focusing your energy on diving deeper in your data and discovering the extent of its modelling capabilities.

After spending only a week familiarizing myself with the interface, I felt confident I could perform my normal tasks with ease and with better hyperparameter tuning and more comprehensive model evaluations.

If you are wondering if these simplifications limit the customization of models, think again. For most uses, the customization options available match the level programming provides by utilizing features such as drop-down menus and editable text boxes – eliminating unnecessary mental overhead. 

Open Source Code node as a supervised learning node and example code in code editor

Still itching to program?

You have options: the SAS Code node and the Open Source Code node (available for use with R and Python). Both nodes can be used in any part of the pipeline, including preprocessing, supervised learning, and post-processing.

For example, you may have preprocessing code for extra messy data already written in R. All you need to do is add the Open Source Code node, insert your code, and update the variables to match the provided macros. Or, maybe you want to try the Deep Learning toolkit and CAS action? Drop a SAS Code node into your pipeline, add your code, and you are good to go!

Little to no programming experience or not quite a machine learning expert?

The drag and drop interface, wide selection of templates, and the extensive evaluations allow for almost anyone to produce professional-level results in a matter of minutes. While using SAS VDMML might not require expert-level knowledge on machine learning, important projects should have an expert review the approach and results. 

Example of creating a new pipeline using an advanced template

A sample of options for preprocessing and supervised learning.

The days of spending weeks programming scripts for feature extraction, fine-tuning models, and evaluating your model are over!

For example, let’s say I’m attempting to impute some variables using R. First, I might store the names of  each columns separately based on the type of imputation I want to perform on it. Then, I could create the code for each type of imputation. If I’m only attempting 2 different types of imputation, I will most likely need less than 10 lines of code. Not much, right?

But, I will also need to test and verify that each variable has been imputed correctly.

Instead, the same task in SAS VDMML would just require you to drop an Imputation node into the pipeline and select via a drop-down menu how to impute the variables – no time wasted.

Additionally, you can quickly compare a variety of supervised learning methods as well as test out the same model with different pre-processing methods using the automatic evaluations provided.  

Looking to save even more time?

You can use the autotuning feature to select the best set of hyperparameters for your model by turning it on in your supervised learning node of choice and hitting run. 

Turn on the autotuning feature inside your supervised learning node, then adjust ranges for the hyperparameters.

After the run is complete, view the supervised learning model’s results to see the best configuration of parameters as determined by autotuning. 

Example of the results after using autotuning for a decision tree.

All of this can be accomplished with a few clicks, which eliminates the hours spent debugging scripts and connecting the steps in your workflow.

You’ve spent the last hour transforming and creating features

Fortunately, SAS VDMML only allows you to view results if the pipeline is up-to-date, which ensures that all changes are accounted for. Now, instead of checking and checking again that I passed the right data to the right functions, I can immediately know that my small tweak had no effect on the results. *sigh* OR on a brighter note, that the drastic improvement is not a fluke!  

Updating the Feature Extraction node resets all child nodes below – ensuring that the pipeline stays up-to-date.