Neural Collaborative Filtering (NCF)
Matrix factorization scores a user-item pair with a dot product — a fixed, linear way of combining two embeddings. Neural collaborative filtering starts from a question about that choice: if the interaction function is just another thing a network could learn, why prescribe it?
The three models
- GMF — the element-wise product of the user and item embeddings, but with learned output weights instead of a plain sum, so classic matrix factorization becomes a special case.
- MLP — the two embeddings concatenated and pushed through hidden layers, able to represent interactions an inner product cannot express.
- NeuMF — both paths trained with their own embeddings and fused at the final layer, so the linear and non-linear views contribute independently.
What the framing changed
NCF recast recommendation from rating prediction to implicit-feedback classification: observed interactions are positives, items drawn from the catalogue are negatives, and the loss is binary cross-entropy. That setup — along with the leave-one-out, 99-negative MovieLens protocol published with it — became the default experimental template for years of papers, which shaped the field as much as the architecture did.
The honest caveat
A careful 2019 reproducibility study found well-tuned matrix factorization matching or beating NCF on its own benchmarks, and the sampled evaluation it popularised is now known to distort top-k metrics. The idea remains the right mental model for learned interaction functions, and the MLP path still earns its place in ranking stages — treat the reported gains with suspicion, not the concept.
Neural collaborative filtering (NCF) replaces matrix factorization’s fixed dot product with a learned interaction function, combining a generalized factorization path (GMF) and a multi-layer path (MLP), fused as NeuMF. It is trained as implicit-feedback classification with sampled negatives rather than rating regression, which is how most recommenders are framed today. Later reproducibility work showed well-tuned factorization matches it, so the framing outlived the reported gains.
Older recommendation software followed a set recipe for deciding how well a person and a product match: line up their descriptions, multiply, add. Sensible, but fixed.
Neural collaborative filtering asked why that recipe should be handed down at all, and let the software learn its own way of matching people to things from the record of what everyone has clicked and bought. The idea reshaped the field — it also reset how these systems are tested, treating every past interaction as a yes and the untouched catalogue as a rough no. Later checks showed the old recipe, tuned properly, was never far behind.