Shapley value

So a second AI just to watch the first one.

A common method involves things like feature attribution mapping or calculating what are called SHAP values.

SHAP values.

What is that?

SHAP stands for Shapley Additive Explanations.

That sounds incredibly complicated.

It's a concept borrowed from cooperative game theory.

I don't show this data, but basically PBC and women have a particular advantage with our model.

And with Shapley analysis, we can identify what are the key features that are driving that prediction.

So reassuringly, we see that the key features that comprise the MEL 3.0 are the key drivers of risk.

So this is very reassuring because those are also found in the MEL 3.0.

This is an integration of our model into Epic, which is our electronic medical record system, really courtesy of our data engineer who was able to implement the solution enable automated pull of the data to then feed the dashboard and provide individualized predictions with our Dynamel score.

And you can see how the patient does dynamically over time, as well as relative to various other patients on the wait list.

So you can see that as well, their survival benefit and the Shapley analysis showing which are the key factors that are driving that patient's risk at that particular point in time.

I should also mention an excellent model that has been recently published by a group from Stanford, wherein they developed a light GBM model, so a gradient boosting model that had an excellent AUC to predict futile procurements of DCD organs in liver transplantation.

Right.

So did the striker contribute 40 percent of the win? Did the goalie contribute 20 percent? SHAP does the exact same thing.

but with data features instead of players.

So if an AI denies your loan, SHA looks at all your features, your income, your credit score, your zip code, and calculates exactly how much each individual feature contributed to the final denial.

I mean, with all the MLPs and foundation models, I mean, most of the time they're treated as black box.

I hear much more of SHAP than of LIME.

Like TAP-PFN and TAP-ECL have SHAP adaptations.

So you can compute SHAP values.

You could, of course, throw permutation importances.

Yeah, other than that...

And so you're like, "Okay, this source is worth close to a cent," feels like, right? That's intuitive.

Now, the formalization of this kind of an intuition is the, the, the core framework we use, it's called Shapley values.

Uh, it's a-

Yes.

Let's go.

What's a Shapley value?

It's a-

Music to my ears