Hi guys. Thank you very much for the great course.
I have a doubt on how to adapt the Collaborative Filtering technique to product recommendation. I am trying this approach for the Santander Product Recommendation in Kaggle.
On Movielens, we have ratings in a sparse Matrix of users x movies. The ratings approach spares a lot of memmory space once we have to train only using the movies rated (a small percentage of the matrix).
But for product recommendation I have only 0/1 target variables. A costumer do or do not have a product in a point in time, so to train the model I have to train with each combination of customer x product, once a NN with only 1´s targets wont train and does not make sense. The problem is that the matrix gets huge and I eventually get a memmory overflow.
I tested the approach with only 1 month of data and it is promising. I could get a good score on the leaderboard. But I would like to use all the periods of data to test if I can improve the results. The problem is that the matrixes get really big.
Any suggestions on how to use the Collaborative approach for have/doesnt have or bought/did not buy product recommendations so I can estimate the products a customer is more likely to buy?
Bst regards,
Fabio.