Research map / Architectures and efficiency / RNNs and normalization
Interpretable Mesomorphic Networks for Tabular Data
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 Interpretable Mesomorphic Networks (IMN), where a deep MLP hypernetwork generates per-instance linear models for tabular classification/regression, giving explanations by design while matching black-box accuracy.
Problem
Neural networks on tabular data are not explainable by design; existing explainable-by-design deep models (e.g., TabNet, NAM) trade off accuracy or are slow, and post-hoc explainers require per-instance surrogate fitting.
Approach
Train a deep MLP hypernetwork that takes an instance x and outputs the weights of a linear model applied to that same x, end-to-end with the loss, so predictions are linear in features per instance and explanations are the weight-times-feature terms, with one forward pass and no per-point surrogate.
Key results
- IMN achieves the best average rank on the AutoML benchmark AUROC against decision trees and logistic regression, with statistically significant differences.
- IMN's AUROC is statistically indistinguishable from TabResNet and CatBoost, while outperforming TabNet and NAM.
- IMN ties for most wins (7 of 40 metric-dataset cells) on XAI-Bench and is 64x faster in inference than TabNet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the black-box ResNet classification head with instance-specific linear weights generated by a hypernetwork | ResNet50 on CIFAR-10 · CIFAR-10 validation | accuracy: 87.49 ± 1.73 vs 88.76 ± 1.50 | No clear effect |
| Use IMN rather than TabNet for instance-level explanations | tabular explainability inference · Credit-g, Adult, Christine | inference time: IMN: 0.01, 0.02, 0.02; TabNet: 0.11, 1.30, 0.43 | Helped |
| Use IMN rather than TabNet for explainability | XAI-Bench tabular datasets · Gaussian Linear, Gaussian Non-Linear, Gaussian Piecewise | total wins across interpretability metrics: 7 vs 0 | Helped |
| Use IMN rather than Kernel SHAP for explanations | XAI-Bench tabular datasets · Gaussian Linear, Gaussian Non-Linear, Gaussian Piecewise | total wins across interpretability metrics: 7 vs 7 | No clear effect |
| Use IMN rather than TabResNet as a classifier | AutoML benchmark · AutoML benchmark datasets | AUROC | No clear effect |
| Use IMN rather than CatBoost as a classifier | AutoML benchmark · AutoML benchmark datasets | AUROC | No clear effect |
| Use IMN rather than TabNet as a classifier | AutoML benchmark · AutoML benchmark | AUROC | Helped |
| Use IMN rather than NAM as a classifier | AutoML benchmark · AutoML benchmark | AUROC | Helped |
| Generate local linear hyperplanes with IMN | Half-moon toy task · 10 to 200 neighboring examples | accuracy: 0.84, 0.82, 0.78, 0.77, 0.77 | Mixed |
| Train IMN for mushroom edibility prediction | IMN · Mushroom edibility test | AUROC: 1 | Helped |
Surprising findings
- IMN's inference was reported as 11-65x faster than TabNet and 1710-11400x faster than SHAP on TabResNet, rather than just comparable accuracy.
- On the Adult dataset the IMN feature ranking showed accuracy drops monotonic with importance, while CatBoost, TabNet, SHAP and decision trees did not.
Limitations
- Per-instance models are restricted to linear functions; non-linear interpretable models are left to future work.
- Explanation quality drops in the presence of feature correlations, similar to other explainers.
- Explainability experiments use only the synthetic XAI-Bench Gaussian datasets and an image appendix with limited comparison.
Experimental setup
Datasets: AutoML benchmark (35 classification datasets) XAI-Bench Gaussian Linear XAI-Bench Gaussian Non-Linear XAI-Bench Gaussian Piecewise Adult Census Income Mushroom CIFAR-10
Benchmarks: AutoML benchmark XAI-Bench
Metrics: AUROC faithfulness monotonicity ROAR infidelity Shapley correlation inference time accuracy drop on feature removal
Baselines: Logistic Regression Decision Tree Random Forest CatBoost TabNet NAM TabResNet SHAP LIME Kernel SHAP L2X MAPLE
Builds on: TabResNet TabNet Neural Additive Models Optuna
Closest papers by meaning
- InterpreTabNet: Distilling Predictive Signals from Tabular Data by Salient Feature Interpretation (ICML 2024 · similarity 0.86)
- TabM: Advancing tabular deep learning with parameter-efficient ensembling (ICLR 2025 · similarity 0.86)
- Neural Reasoning Networks: Efficient Interpretable Neural Networks with Automatic Textual Explanations (AAAI 2025 · similarity 0.85)
- Hybrid Autoencoders for Tabular Data: Leveraging Model-Based Augmentation in Low-Label Settings (NeurIPS 2025 · similarity 0.85)
- Making Pre-trained Language Models Great on Tabular Prediction (ICLR 2024 · similarity 0.85)
- Interpretable Deep Clustering for Tabular Data (ICML 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.