topic Re: How can I get lime plot for a specific observation in sas viya 3.5? in SAS Data Science

How can I get lime plot for a specific observation in sas viya 3.5?

Ullsokk — Mon, 03 Apr 2023 08:59:01 GMT

How do I get a LIME explanation for a specific observation is sas viya 3.5? I want to be able to explain why a specific customer has a high probability value on a model.

Re: How can I get lime plot for a specific observation in sas viya 3.5?

sbxkoenk — Mon, 03 Apr 2023 10:06:30 GMT

@Ullsokk wrote:

How do I get a LIME explanation for a specific observation is sas viya 3.5? I want to be able to explain why a specific customer has a high probability value on a model.

I moved this topic to "Data Science board".

A LIME explanation for one specific observation ??
To explain your high probability value you need to estimate Shapley values!

Run the shapleyExplainer action in PROC CAS (HyperSHAP and KernelSHAP methods are available).

In this blog they explain how to do : Parallel Shapley value estimation
https://blogs.sas.com/content/subconsciousmusings/2021/07/06/tips-for-parallel-processing-in-casl/

(it might be worthwhile to do it in parallel because it takes a lot of resources

, so if you do not have a lot of resources --> then it takes time)

Good luck,
koen

Re: How can I get lime plot for a specific observation in sas viya 3.5?

sbxkoenk — Mon, 03 Apr 2023 10:12:45 GMT

On top of previous response ...

Subject: interpreting complex (machine learning) models

partial dependency plots
variable importance plots / Variable Importance Rankings
LIME (Local Interpretable Model Agnostic Explanations)
ICE (Individual Conditional Expectation)
SHAPLEY values

Some resources (#5):

LIME and ICE in SAS Viya 3.4: Interpreting Machine Learning Models (article)
https://communities.sas.com/t5/SAS-Communities-Library/LIME-and-ICE-in-SAS-Viya-3-4-Interpreting-Machine-Learning/ta-p/510915
See also REFERENCES and ADDITIONAL RESOURCES in the bottom of article!
Paper SAS1950-2018
Interpreting Black-Box Machine Learning Models Using Partial Dependence and Individual Conditional Expectation Plots
Ray Wright, SAS Institute Inc.
https://www.sas.com/content/dam/SAS/support/en/sas-global-forum-proceedings/2018/1950-2018.pdf
SAS Tutorial | Interpreting Machine Learning Models in SAS (video van 41m10s)
https://www.youtube.com/watch?v=6LcyVSLwVck
Communicating Analytical Results and Interpreting Machine Learning Models with SAS Viya (video van 44m11s)
https://www.youtube.com/watch?v=Ql5oB-CDFGU
Model Interpretability Plots - SAS Help Center
https://go.documentation.sas.com/?docsetId=vdmmladvug&docsetTarget=n0dqb7uyhlqogvn16xnengpro0fc.htm&docsetVersion=8.4&locale=en

Enjoy!

Koen

Re: How can I get lime plot for a specific observation in sas viya 3.5?

xinhunt — Thu, 06 Apr 2023 18:52:31 GMT

As Koen has explained in an earlier reply, Shapley values may be more appropriate for your use case because Shapley values give you the importance of each input variable for a prediction. LIME values tells you "how much would I expect the prediction to change if I change one/some of the inputs near this observation" which is not exactly the same as feature importance, but could be useful for different scenarios.

To calculate Shapley values or LIME values, you can use the explainModel action set. Here are some links for code examples -- all links are for Viya3.5:

(1) Example for calculating Shapley values using ShapleyExplainer (HyperSHAP, preferred method for Shapley value estimation)

(2) Example for calculating LIME values

(3) Example for calculating Shapley values using linearExplainer (KernelSHAP, can be less accurate than HyperSHAP, but in case you need an alternative method):