Background
I'm building a binary classification model for a pair match problem using CNN, e.g. whether person A1 likes product B1 or not. Model input features are sequence features (text descriptions) of the person and the product. The model accuracy is around 78%. So for a new person, the model can predict the probability whether he likes each product in our dataset.
Problem
The model is good if we know nothing about the person. However, in the real scenario, we already know the new person likes one or two products. We want to predict whether he likes other products. Is there any way to incorporate this prior information to improve the model?
My thought
A simple method would just retrain the model, giving the new person's pair higher sample weight. But we can't do this for each new person.
Any suggestion would be appreciated. Thanks
