Collaborative Filtering

Collaborative filtering uses the wisdom of ‘similar’ crowd in order to find recommendations for user A based on interests of a similar user B.

This method has popularity bias. Hence:

• It needs to be re-trained every few minutes (or hours) so that the model stays up to date with what’s popular.
• It cannot be used for use-cases like recommendations in a pharmacy.

Why it matters: At scale, most products cannot be targeted. For example, a specific SUV might be targeted towards off-roading enthusiasts but an urban professional who had a recent hike might also buy it not because the product was built for him but because they can.

Collaborative filtering typically requires matrix factorization.