Recommender System: how to treat different events

Question

I'm trying to build recommender based on user history from e-commerce. There are two(potentially more) types of events: purchase and view. Is it okay to sum up number of purchases and views for a given item(with purchase and view having different weights)? Or I`ll just mix up user intent this way?

Answer 1

I assume you're using an implicit feedback recommender approach like ALS. Otherwise, summing data points generally won't make sense, such as if you're feeding it to a recommender that expects ratings.

The input to implicit ALS is, conceptually, weighted user-item pairs. Therefore it makes sense to perhaps use a sum of user-item clicks as the weight. Summing makes sense.

However, does a purchase and view seem to carry the same weight? obviously not. A purchase is a much stronger association and should be weighted accordingly.

As to how much, I'd suggest weighting purchases simply by price, to start. Then weight clicks by price times purchase-to-click ratio. A $10 item purchase is weight 10; if 1 in 200 clicks results in a purchase, then weight a click 0.2.

This is crude but probably about as close as anything for capturing this info in the context of ALS.