Research map / Architectures and efficiency / RNNs and normalization
Neural Reasoning Networks: Efficient Interpretable Neural Networks with Automatic Textual Explanations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Neural Reasoning Networks (NRNs), a neuro-symbolic architecture of weighted Lukasiewicz logic neurons, and R-NRN, a bandit-guided training algorithm for tabular classification that yields compact natural-language logical explanations.
Problem
Neural networks for tabular classification are hard to interpret, and existing interpretable models (NAM, CoFrNet) give only feature-importance explanations that grow with feature count and are slow or large to train.
Approach
Networks of logical And/Or neurons based on modified weighted Lukasiewicz logic, whose structure is learned by a Bayesian UCB multi-armed bandit alongside gradient descent, with explanations generated by traversing the network and simplifying it into a single conjunction of rules.
Key results
- R-NRN average normalized test AUC 0.980 vs RF 1.000 and XGB 0.993, statistically similar to RF, XGB, GBT, NAM, CFN, DAN.
- R-NRN uses about 1K parameters on average versus 861K for NAM and 37K for CFN, and trains 43% faster than NAM and CFN on average.
- R-NRN explanations are 31% smaller (14.8 vs 21.3) than NAM/RF feature-importance explanations, with higher single-deletion Pearson correlation than NAM and RF.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use R-NRN instead of MLP | tabular classification · 22-dataset tabular classification benchmark | normalized test AUC: 0.865 -> 0.980 | Helped |
| Use R-NRN instead of BRCG | tabular classification · 22-dataset tabular classification benchmark | normalized test AUC: +7% | Helped |
| Use R-NRN instead of DIF | tabular classification · 22-dataset tabular classification benchmark | normalized test AUC: +16% | Helped |
| Use R-NRN instead of FTT | tabular classification · 22-dataset tabular classification benchmark | normalized test AUC: +28% | Helped |
| Use R-NRN instead of RF, XGB, or GBT | tabular classification · 22-dataset tabular classification benchmark | normalized test AUC | No clear effect |
| Use R-NRN instead of NAM or CFN | tabular classification · 22-dataset tabular classification benchmark | normalized test AUC | No clear effect |
| Generate explanations with R-NRN rather than feature-importance approaches | sample-level model explanations · 22-dataset tabular classification benchmark | explanation size: 31% smaller | Helped |
| Use R-NRN rather than NAM or RF explanations | sample-level model explanations · 22-dataset tabular classification benchmark | Single Deletion Spearman correlation: 0.80 vs 0.74; 0.80 vs 0.73 | Helped |
| Use R-NRN rather than SHAP(RF) explanations | sample-level model explanations · 22-dataset tabular classification benchmark | Single Deletion Pearson correlation: 0.83 vs 0.91 | Hurt |
| Generate explanations with R-NRN rather than SHAP(RF) | sample-level model explanations · 22-dataset tabular classification benchmark | feature importance computation time: 96% faster | Helped |
Surprising findings
- Neural tabular models (NAM, CFN, DAN) are competitive with tree ensembles under AUC and careful tuning, contrary to prior reports.
- SHAP on RF gives better single-deletion Pearson correlation (0.91) than R-NRN (0.83).
Limitations
- Evaluated only on tabular classification; regression, computer vision and NLP not tested.
- Structure modification flexibility is limited by the initialized architecture.
- No user study evaluating the usefulness of explanations.
Experimental setup
Datasets: OpenML-CC18-style tabular benchmark (22 classification datasets, OpenML IDs 44089-44162)
Benchmarks: 22 OpenML tabular classification datasets
Metrics: ROC AUC (micro-averaged) training time parameter count explanation size single deletion Spearman/Pearson correlation
Baselines: BRCG DIF NAM CoFrNet FT-Transformer DANet MLP Random Forest XGBoost Gradient Boosted Trees
Builds on: Logical Neural Networks Neural Additive Models CoFrNet Optuna TPESampler
Closest papers by meaning
- Learning Accurate and Interpretable Decision Rule Sets from Neural Networks (AAAI 2021 · similarity 0.87)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.87)
- Systematic Relational Reasoning With Epistemic Graph Neural Networks (ICLR 2025 · similarity 0.86)
- Explaining, Fast and Slow: Abstraction and Refinement of Provable Explanations (ICML 2025 · similarity 0.86)
- Interpretable Mesomorphic Networks for Tabular Data (NeurIPS 2024 · similarity 0.85)
- GraphTrail: Translating GNN Predictions into Human-Interpretable Logical Rules (NeurIPS 2024 · similarity 0.85)
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.