Shapley Values for Machine Learning 

Hi Louis, I am glad I could help. Good luck for your internship!

Thank you so much! I really appreciate the support :)

@@cauchyschwarz3295 This is a bit confusing! But no, you will only have one model (the one you want to explain). You marginalise the prediction function (I.e. model) over the features that are not in set S. You do not have to retrain a new model by excluding the features that are not in S.

I think there's a subtlety here in the application. In the theoretical model the goal is to calculate a fair payout for an individual feature based on how much it contributes to building a good model. Good models are measured by how well they predict things. So, we would want to think about how much the feature contributes to a reduction in model error, and you would want to train a model with and without a given feature to figure this out. (But instead we use other methods for determining which features are good that are easier to compute.) In the use case in this video (and what is the norm in ML), they are using Shapley to explain how a feature contributed to the model prediction, irregardless of how good the model actually is. This is helpful because Shapley still has desireable properties like additivity in this use case. If you trained a new model without this feature, it wouldn't answer the question of how much the feature contributed to the prediction value in the old model.

I don't see the mistake. At what time period do I introduce (Val S U{1})? {1} referes to a coalition of feature 1 i.e. x1 and not the feature's values

Thanks! If the continuous variable is in the coalition, then we use the actual value for that instance (i.e. the person's actual age). If the continuous variable is not in the coalition, then we integrate over the values of the variable w.r.t. to the probability of the values. However, in practice, we will not know the probability distribution of a variable. So we will have to randomly sample different values for the variable from our dataset. We do this a bunch of times so we end up approximating the distribution. I hope that makes sense? There is a lot of statistical theory that underlies this explanation!

@@adataodyssey Ahh I see, got it. But out of curiosity, Can you give me the reference of those statistical theory?

@@chrisleenatra Unfortunatley, I don't have any specific references. I'm using the knowledge from back in my undergrad. If you want to understand take a look at "stochastic calculus"

Sure, Avijeet! You can find all the videos in this playlist: SHAP ru-vid.com/group/PLqDyyww9y-1SJgMw92x90qPYpHgahDLIK

You can but they will cancel out when you subtract val(S) from val(S U {i})

Aaaaah of course, that makes sense! Thanks, you helped me a lot - not only with this comment but the entire video series:)

@@NeverHadMakingsOfAVarsityAthle No problem Matthias! I'm glad I could help :)

Thank you Erica!

Yes, that's correct. The efficiency property of Shapley values doesn't have anything to do with computational efficiency. It just means that if you add the Shapley values of an instance and the mean prediction across all instances, you will get the prediction for that instance.