A few months back, at work, a couple of kids in my team taught me this concept called “SHAP“. I won’t go into the technical details here (or maybe I will later on in this post), but it is basically an algo that helps us explain a machine learning model.
It was one of those concepts that I found absolutely mind-blowing, to the extent that after these guys taught this concept to me, it became the proverbial hammer, and I started looking for “nails” all around the company. I’m pretty sure I’ve abused it (SHAP I mean).
Most of the documentation of SHAP is not very good, as you might expect about something that is very deeply technical. So maybe I’ll give a brief intro here. Or maybe not – it’s been a few months since I started using and abusing it, and so I’ve forgotten the maths.
In any case, this is one of those concepts that made me incredibly happy on the day I learnt about it. Basically, to put it “in brief”, what you essentially do is to zero out an explanatory variable, and see what the model predicts with the rest of the variables. The difference between this and the actual model output, approximately speaking, is the contribution of this explanatory variable to this particular prediction.
The beauty of SHAP is that you can calculate the value for hundreds of explanatory variables and millions of observations in fairly quick time. And that’s what’s led me to use and abuse it.
In any case, I was reading something about American sport recently, and I realised that SHAP is almost exactly identical (in concept, though not in maths) to Wins Above Replacement.
WAR works the same way – a player is replaced by a hypothetical “average similar player” (the replacement), and the model calculates how much the team would have won in that case. A player’s WAR is thus the difference between the “actuals” (what the team has actually won) and the hypothetical if this particular player had been replaced by the average replacement.
This, if you think about it, is exactly similar to zeroing out the idiosyncrasies of a particular player. So – let’s say you had a machine learning model where you had to predict wins based on certain sets of features of each player (think of the features they put on those otherwise horrible spider charts when comparing footballers).
You build this model. And then to find out the contribution of a particular player, you get rid of all of this person’s features (or replace it with “average” for all data points). And then look at the prediction and how different it is from the “actual prediction”. Depending on how you look at it, it can either be SHAP or WAR.
In other words, the two concepts are pretty much exactly the same!