1
$\begingroup$

I know SHAP (or shapley) values are the contribution of each input variable to the model prediction. Adding the base values to the sum of all SHAP values gives you the model prediction for any data point. I dont really understand what the base-value here actually imply.

Imagine a binary classification model predicts absence (0) or presence (1) of a crop, and I train the model with equal data points of absence and presence, then the base-prediction is 0.5. And suppose the model inputs are various environmental variables. If, at a pixel, the SHAP values of all environmental variables sum to zero, the model prediction is 0.5, or there is 50% chance that a crop is present. My question is what contributes to that 50% chance if all variables eventually contributed zero?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
1
$\begingroup$

In SHAP (SHapley Additive exPlanations), the base prediction value simply represents the expected value or the average prediction of your model across your entire background dataset. It serves as the model's "default guess" before it considers any specific features of a particular data point. In your example of a perfectly balanced dataset predicting the absence or presence of a crop, the base prediction is exactly 0.5.

To answer your question about what contributes to that 50% chance when the variables seemingly contribute zero, that 50% comes entirely from the overall average of the training population, not from the specific variables at that given location. When the SHAP values of all environmental variables at a pixel sum to zero, resulting in a 0.5 prediction, it means that the specific combination of environmental factors provides no net new information to push the prediction away from the baseline probability. The positive and negative influences of those specific variables perfectly cancel each other out, leaving you relying solely on the population's baseline average

Improve this answer
$\endgroup$

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.