Can MLPs learn to match dot product similarity in practice?

The Neural Collaborative Filtering paper popularized replacing the dot product with a learned MLP for combining user and item embeddings. The justification was theoretical: an MLP is a universal function approximator, so it can in principle learn any similarity function — including dot product — and presumably better ones. Rendle et al.'s revisit shows this argument fails empirically and operationally.

Empirically, with careful hyperparameter selection, a properly configured dot product baseline substantially outperforms the MLP. Even more pointedly, learning a dot product through an MLP requires a large model capacity and a lot of training data — the universal approximation guarantee is asymptotic, and finite-data inductive bias matters more than expressiveness. The MLP is too flexible for the task; its inductive bias points away from the simple geometric similarity that actually fits the data.

Operationally, dot products allow maximum-inner-product search over precomputed item embeddings, which is fast enough for real-time serving over millions of items. MLP similarities require a forward pass per item per query — they cannot be precomputed. So even if MLPs were marginally more accurate, they would be unaffordable in production.

The takeaway: an inductive bias that matches the geometry of the problem (dot product) wins over an expressive parameterization that has to learn the geometry from scratch.