1

I have a CatboostRanker (with groups) model trained on a dataset of ~1m rows of ~300 features. Three such features are ~90% invalid (have a special category denoting such). My data is time series based, so the invalid data of these features is all but the most recent data.

For some reason, these three features are amongst the top 5 most important according to their shapley values (the absolute sum of all shap values for a given feature). When looking at the individual shap values for each individual object, ALL of them are positive, meaning they all contribute positively to the target variable of binary [0,1].

I don't understand why the 90% of invalid values for these features all carry with them a positive shapley value, since the invalid category theoretically confers no helpful information. I've read several explanations of shap values, and understand their mathematical basis, but still no closer to an answer. Any suggestions would be much appreciated.

On the other hand, Catboost's own permutation-based get_feature_importance method ranks these variables lowly which seems far more intuitive.

Method: I used shap.TreeExplainer on the trained model, then extract the shap values from the explainer by passing a catboost pool containing my training data.

Improve this question

1 Answer 1

Reset to default
0

SHAP does not determine if your features are valid or not. It determines the importance of those features to your model's output. I had the same confusion when implementing it for some use cases.

If you go back to the game theory origin of SHAP, you can understand it as this : SHAP does not determine if the player (your feature) is a good player (valid feature) but it determines how much it contributes to the game (your output). If you want to gain understanding on the validity of your features, you should see the model's performance as well (your metrics).

Improve this answer
Sign up to request clarification or add additional context in comments.

2 Comments

Thanks Laassairi. My question in this case would be, why would a feature with 90% missing/invalid values contribute highly to the outputs (the predictions)? For those 90% of objects with missing values in these features, why would those objects' shap values for those features be high? Because if they're missing/invalid, these feature/values should not be contributing towards the outputs/predictions at all.
Catboost has 3 modes of interpretting missing values (min, max and forbidden) the default being min. So 90% of your missing values aren't actually "None" once they're being processed by the model. Furthermore, a feature's contribution to the output isn't proportional to the % of its missing values. Imaging you have Y = 1, 2, 3, 4 ... to N. And X1= np.random(size=N), X2=np.random(size=N), X3 = None, None.., N/2, N/2+1, ... to N. Where half of X3 is missing. In this use :case X3 would likely contribute more to the model's output (Y) than X1 and X2.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.