Research map / Architectures and efficiency / RNNs and normalization
Polyhedron Attention Module: Learning Adaptive-order Interactions
Tan Zhu, Fei Dou, Xinyu Wang, Jin Lu, Jinbo Bi
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 |
|---|---|---|---|
| Remove overlapping polyhedrons (hard split) (ablation) | PAM · Criteo, Avazu, UK Biobank | AUC / R2 | Improved |
| Add adaptive interaction order via upper bound 2Ui (ablation) | PAM · Criteo, Avazu, UK Biobank | AUC / R2 | Improved |
| Remove ||Wi|| from the denominator of Eq. 7 (ablation) | PAM · Criteo, Avazu, UK Biobank | AUC / R2 | Improved |
| Vary initial value of Ui from 1 to 3 | PAM · Criteo, Avazu, UK Biobank | AUC / R2 | No clear effect |
| Increase oblique tree depth D | PAM · Criteo, Avazu, UK Biobank | AUC / R2 | Mixed |
| Use PAM instead of second-best baseline | PAM · Criteo | AUC (per mille): 1.4 | Improved |
| Use PAM instead of second-best baseline | PAM · Avazu | AUC (per mille): 0.6 | Improved |
| Use PAM instead of second-best baseline | PAM · UK Biobank | R2: 2.2 | Improved |
| Use PAM instead of DNN, training cost on Criteo | PAM · Criteo | training sec/batch: 0.046 -> 0.102 | Hurt |
| Use PAM instead of DCN-V2, inference cost on UK Biobank | PAM · UK Biobank | inference sec/batch: 0.0067 -> 0.0018 | Improved |
| Use PAM instead of AOANet, training cost on Avazu | PAM · Avazu | training sec/batch: 0.338 -> 0.090 | Improved |
| Use PAM instead of DCN, inference cost on Criteo | PAM · Criteo | inference sec/batch: 0.0114 -> 0.0330 | Hurt |
Closest papers by meaning
- Provable Explanations for Any-Order Neural Additive Models (NeurIPS 2026 · similarity 0.83)
- SmoothHess: ReLU Network Feature Interactions via Stein's Lemma (NeurIPS 2023 · similarity 0.83)
- Poly-attention: a general scheme for higher-order self-attention (ICLR 2026 · similarity 0.83)
- Infinite-Dimensional Feature Interaction (NeurIPS 2024 · similarity 0.81)
- Degrees of Freedom for Linear Attention: Distilling Softmax Attention with Optimal Feature Efficiency (NeurIPS 2025 · similarity 0.81)
- Three Mechanisms of Feature Learning in a Linear Network (ICLR 2025 · similarity 0.81)
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.
