Building a binary classifier on uncertain 0's When building models to predict probability of sales etc. Its intuitive to select customers who already have bought the product as training data for class 1 and customers who does not have the product as class 0.
But, many of the customers who does not have that product may have never been introduced to it, and would have bought it if they were. So they don't really belong in the 0 class in the training data.
Doe anyone have any thoughts/reflections on this? Whats the consequence?, can we do something to mitigate etc.
 A: In binary classification, if you can't detect a given input for class 1, it is in class 0. Hence, you just need to consider is the input in class 1 which you have their properties and rich training data. Therefore, if an input data does not classify as a member of class 1, it would be in class 0.
A: That's the limitation of binary. It's either be 1 or 0 no other way around it. If you can find a way to introduce some extra measurements relevant to the group you are suggesting, then you can make it a multi class case and build you new model on that.
A: You could regard the customers (1) of which you know that they have bought the product, assume that other very similar customers (0->1) also have already bought the product and that only the rest of the customers (0) has not bought it, yet.
You would need to define "very similar". E.g. it could be cosine similarity between a customer (0->1) and some customer (1) > 90. Or e.g. "very similar" could be customer (0->1) is in a dense cluster with a majority of customers (1) who you know have bought the product.  
But you cannot evaluate the correctness of this approach because the information that a customer has not bought the product, yet, is missing. 
