Clicks Don't Lie, They Just Whisper

A team switches their recommender's training signal from 1-5 star ratings to raw clickstream and dwell-time logs, then fits a matrix factorization model that decomposes the user-item matrix into low-dimensional latent factor vectors for users and items. Compared to the star ratings, what is the key modeling challenge introduced by using clicks and dwell time?

Clickstream data eliminates matrix sparsity entirely

There are no explicit negative signals, since a missing interaction could mean disinterest or simply that the user never saw the item

Matrix factorization cannot be applied to any data that isn't a 1-5 scale

Latent factors can only be learned from explicit numeric ratings, never from behavioral logs