Research map / Learning theory and algorithms / Matrix and tensor methods
RaFM: Rank-Aware Factorization Machines
Xiaoshuang Chen, Yin Zheng, Jiaxing Wang, Wenye Ma, Junzhou Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use feature-specific interaction ranks selected by feature occurrence frequency | FM · ML 10M, ML 20M, AMovie, Frappe, ML Tag, Avazu, Criteo | square loss or log loss: 0.8016 -> 0.7870; 0.8002 -> 0.7807; 1.0203 -> 0.9986; 0.1702 -> 0.1447; 0.2538 -> 0.2387; 0.3817 -> 0.3801; 0.4471 -> 0. | Improved |
| use feature-specific interaction ranks selected by feature occurrence frequency | FM · Frappe, ML Tag, Avazu, Criteo | AUC: 0.9771 -> 0.9811; 0.9503 -> 0.9526; 0.7761 -> 0.7826; 0.8030 -> 0.8060 | Improved |
| combine ranks through feature-specific embeddings rather than DiFacto's truncation and parameter sharing | FM · ML 10M, ML 20M, AMovie | square loss: 0.7950 -> 0.7870; 0.7948 -> 0.7807; 1.0268 -> 0.9986 | Improved |
| combine ranks through feature-specific embeddings rather than DiFacto's truncation and parameter sharing | FM · Frappe, ML Tag, Avazu, Criteo | log loss: 0.1711 -> 0.1447; 0.2529 -> 0.2387; 0.3823 -> 0.3801; 0.4470 -> 0.4451 | Improved |
| use RaFM instead of DiFacto | FM · Avazu | AUC: 0.7778 -> 0.7826 | Improved |
| use RaFM instead of DiFacto | FM · Criteo | AUC: 0.8030 -> 0.8060 | Improved |
| use RaFM instead of DiFacto | FM · Frappe | AUC: 0.9771 -> 0.9811 | Improved |
| use RaFM instead of DiFacto | FM · ML Tag | AUC: 0.9450 -> 0.9526 | Improved |
| vary rank set from S1 to S2, S3, or S4 (ablation) | RaFM · ML Tag | log loss: 0.2405; 0.2416; 0.2387; 0.2391 | No clear effect |
| use exponentially increasing rank settings (S2 or S3) rather than arithmetic settings (S4) (ablation) | RaFM · ML Tag | number of parameters: 9.63M; 9.23M; 17.75M | Improved |
| use arithmetic rank settings S4 rather than S1 (ablation) | RaFM · ML Tag | training time: 1.35× | Hurt |
| use a smaller, occurrence-aware rank allocation rather than a fixed rank of 512 | FM · ML Tag | log loss: 0.2538 -> 0.2387 | Improved |
Later papers that compared with this method
1 later accepted paper reports RaFM as a baseline in their results tables; 0 of them report a better number on most of the results they share with it.
- Exploiting Field Dependencies for Learning on Categorical Data (TPAMI 2023 · better on 1 of 2 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Higher-Order Factorization Machines (NeurIPS 2016 · similarity 0.86)
- Accurate and Interpretable Factorization Machines (AAAI 2019 · similarity 0.86)
- Factorization Machines with Regularization for Sparse Feature Interactions (JMLR 2021 · similarity 0.85)
- fastFM: A Library for Factorization Machines (JMLR 2016 · similarity 0.85)
- Attention-over-Attention Field-Aware Factorization Machine (AAAI 2020 · similarity 0.84)
- Boosting Factorization Machines via Saliency-Guided Mixup (TPAMI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
