This is related to this other question but focusing on a particular potential solution. Reading the other question might help in understanding the background of this one, although the other is quite longer.
Is there any way in Keras to force the model to give more attention to a particular feature or combination of feature? Particularly in the case of a Simpson's Paradox, I'd like a specific combination of feature to take precedence even if the model has rarely seen this combination during training.
A toy example: a person who has dark hair and glasses will always like ice cream, even if dark haired people in general dislike ice cream, and glasses-wearing people in general also dislike ice cream. It's just that this specific combination of dark hair + glasses pretty much guarantees the person will like Ice Cream, but unfortunately, the model hasn't seen such a combination as often.
In this example, I'd like to know if there is a way to force the model to consider the combination (hair_color, glasses_or_not) as strongly determinant for the outcome.
I've said it in the other question, but resampling the training data to give more examples of such combinations would be inapplicable in the real-life scenario, as the data is way too complex for this (think 20 features with 5,000 cardinality each, with a myriad of hidden examples similar to the toy illustration above).
