categorial variables about deeplift HOT 2 CLOSED

Hi @misbahch6,

It would depend on three things:
(1) What was the reference a.k.a. baseline that you used?
(2) Which output task are you computing the importance with respect to?
(3) What is the sign of the importance that you observe for a particular example?

Say you used a baseline of all-zeros. Now consider an example where the family history is a 1. The difference-from-baseline for family history for this example is a +1 (because 1-0 = +1). Now let's say that when you compute the scores for output neuron 1, you observe positive importance for family history, whereas if you compute the scores for output neuron 2, you observe a negative importance for family history. This indicates that family history being "1" (rather than zero) contributes positively to output neuron 1, whereas it contributes negatively to output neuron 2. In all, this suggests that a family history of 1 is important for receiving a high score in output neuron 1, whereas a family history of 0 is important for receiving a high score in output neuron 2.

Does that make sense?

Best,
Avanti

from deeplift.

thank you so much. it is super clear response
:)

from deeplift.