Neural Collaborative Filtering
The 2017 framework that replaced matrix factorization's dot product with a neural network — GMF, MLP, and NeuMF.
About
Neural Collaborative Filtering is the paper that dragged collaborative filtering into the deep-learning era. Presented at WWW 2017 by Xiangnan He and colleagues, it argued that matrix factorization's fixed inner product is a needlessly rigid way to model user-item interaction — and proved the point by showing MF is literally a special case of the NCF framework (cripple the network down to element-wise multiplication and you recover MF exactly).
The framework ships three instantiations: GMF (generalized matrix factorization with learned weights), MLP (a multilayer perceptron learning the interaction function from concatenated embeddings), and NeuMF, which fuses the two — GMF for linear signals, MLP for nonlinear ones. Trained with log loss and negative sampling on implicit feedback, NeuMF beat classic MF on MovieLens and Pinterest on hit ratio and NDCG.
NCF's practical reign was always qualified — later work showed well-tuned linear baselines could match it, and it ignores sequence, content, and context — but as the canonical neural CF reference it is still the first thing every recsys course teaches and the baseline every neural CF paper cites.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper-reported (WWW 2017, arXiv 1708.05031): NeuMF improved hit ratio and NDCG over classic matrix factorization baselines on MovieLens 1M and Pinterest datasets
- Widely reproduced; later studies note that tuned linear baselines can match NCF, so treat the headline gains as era-specific rather than absolute
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 9/10
Scored 9/10 (TIER verified, metrics-v1). Evidence: peer-reviewed WWW 2017 paper (arXiv 1708.05031) with author-released reference code (github.com/hexiangnan/neural_collaborative_filtering) plus many community ports; one of the most cited recsys papers of its era and a permanent teaching baseline. Strengths: rigorous framing (MF as a special case), fully open code, enormous replication record. Docked for: (1) headline gains are era-specific — later work shows tuned linear baselines match it; (2) the model family is a starting point, not a production solution. License: open research code.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 10.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 9.0/10
- · WWW 2017 paper (arXiv 1708.05031, He et al., NUS); GMF/MLP/NeuMF framework with MF proven as a special case; author code at github.com/hexiangnan/neural_collaborative_filtering.
- · Paper-reported NeuMF gains on MovieLens 1M and Pinterest; one of the most cited recsys papers of its era, a permanent teaching baseline.
- · Headline gains are era-specific — later work shows tuned linear baselines match it. Maintenance frozen; a starting point, not a production solution.