Can we incorporate product content into C.F?

model: $$ r_{ij} = u + b_i + b_j + [u_i + \sum(x_i)]^T v_j $$

where

• $ u $ is global mean
• $ b_j, b_i $ are user and product biases (would capture popular products)
• $ u_i $ is user loadings on latent factors
• $ x_i $ are user specific features (like average of past items viewed
• $ v_j $ are product loadings on latent factors

Issues:

• we have a user x product matrix with views (clicks). Are views useful to predict?
• we have a user x product matrix with buys (but I think this is one per user)

My thoughts about this C.F by CH

• Maybe we can use views (clicks) as rating. We can normalized all clicks with average for each user and use it as weights for viewed items.