Interpreting a Support Vector Machine Model

[WWD]

In Section 5 of AI and Machine learning, the program introduces Support Vector Machines (SVM).  Within this module, there is an extended discussion of how to interpret the SVM results.  

Assume that the data are perfectly separable by a hyperplane.  Would the two clusters the the “1” cluster (event occurrence) and the “0” cluster (the event did not occur)?

When linear relationships are fit for individual variables, what is the target value for the “non-event” group (all the Y values are 0)?

I know this is a tough question to describe and will be tougher to interpret.  If possible, may I make a phone appointment with an instructor to discuss the problem, please?

Thank you,

Bill Donaldson

[Cynthia_sas]

Hi:
The AI & ML module is composed of 5 separate e-learning classes as shown below:
AI&ML Module composed of 3 sections:
1) Machine Learning certification exam preparation
Class: Machine Learning Using SAS Viya

2) Forecasting and Optimization certification exam preparation
Class: Forecasting Using Model Studio in SAS Viya
Class: Optimization Concepts for Data Science and Artificial Intelligence

3) Natural Language and Computer Vision certification exam preparation
Class: SAS Visual Text Analytics in SAS Viya
Class: Deep Learning Using SAS Software

In order to better help you, can you provide the following information?
=== === ===
Module: AI&ML
Class/Title:
Lesson:
Section:
Video, Demo or Practice Title:
=== === ===

Thanks, in advance,
Cynthia

[WWD]

Hey Cynthia:

All of the questions that I’ve submitted pertain to the module “Machine Learning Specialist” with the title of the curriculum titled “Machine Learning Using SAS Viya”.  One of my questions pertains to SVMs which is chapter 5 of this curriculum.

Another question dealt with neural nets, which is Chapter 4 within this same module.

Thank you,

Bill

[Cynthia_sas]

Hi, Bill:
Thanks for the followup. You don't have to make a new posting to clarify your question, you can either edit your original question or just reply to one of the other posts. I see that Ari (one of the AI&ML instructors) has had a chance to post a response. If you have a more in-depth question for the instructors, you can always send mail to curriculumconsulting@sas.com and just refer to the posting that you have more questions about (you can grab the URL from the browser window).
And then we can send your followup question to the instructors so they can review it when they are not teaching. Sometimes, the forum participants are able to help students and sometimes the students have more follow-up questions than is possible in the forum.
Cynthia

[AriZitin]

Hi!

I was a bit unsure about some of the questions so feel free to clarify if I missed the mark or let me know if you have follow-up questions!

                Assume that the data are perfectly separable by a hyperplane.  Would the two clusters the the “1” cluster (event occurrence) and the “0” cluster (the event did not occur)?

                I am guessing that your question is “Would the two clusters be the “1” cluster and the “0” cluster?

The clustering described in the section on Model Interpretability is just k-means clustering so there is no guarantee that all of the target events will fall into a single cluster, even if the data is separable by a hyperplane. Some of the model interpretability results like Individual Conditional Expectation (ICE) and Local Interpretable Model-Agnostic Explanations (LIME) are calculated based on a single observation to provide explanations for that particular observation (as opposed to Partial Dependence which averages results across the whole dataset). The point of clustering in this situation is to calculate ICE and LIME based on a cluster of observations (and average the results within the clusters) to get a broader picture of how the model works on different portions of the data. This feature has actually been deprecated in newer versions of the software since it tends to be more useful to plot ICE and LIME for individually selected observations than for somewhat arbitrary clusters (the results of a k-means clustering algorithm).

So the clusters here are really just groups of observations that we try to explain together.

                When linear relationships are fit for individual variables, what is the target value for the “non-event” group (all the Y values are 0)?

I might need clarification on this question, in particular I am unsure of the context for when linear relationships are fit for individual variables. In LIME (Local Interpretable Model Agnostic Explanations) a single observation is selected for interpretation and then a linear model is fit ‘locally’ around the observation of interest. This local linear model is fit using all of the input variables for the original model, but the training data for the local model is ‘generated’ by sampling points around the observation of interest and weighting them based on the distance to the observation of interest. The target values for these generated sample points are just the predicted target values from the original uninterpretable model. In this case there is no “non-event” group since the group of generated sample points will include both events and non-events. This video (https://youtu.be/6LcyVSLwVck?t=1601 linked at the time when we start discussing LIME) provides a bit more explanation of how the LIME results are generated.

-Ari Zitin