Neural Collaborative Filtering

Best for: Recommender systems

How it works

Replaces the fixed inner product of matrix factorisation with learned neural layers: users and items get embedding vectors $p_u,q_i$ and a network $f_\theta$ combines them to predict the rating $\hat r_{ui}=\sigma\bigl(f_\theta(p_u,q_i)\bigr)$. The generalised matrix-factorisation variant uses an element-wise product with learnable weights, $f_\theta(p_u,q_i)=h^\top(p_u\odot q_i)$, recovering classic MF as a special case while stacked MLPs allow non-linear user-item interactions. It is trained to minimise rating or relevance loss (MSE or BCE) over observed interactions, typically combined with negative sampling.

Common fields

E-commerce · streaming · ads · marketplaces