Research map / Probabilistic and scientific modeling / Molecules and proteins
Finding Interpretable Class-Specific Patterns through Efficient Neural Search
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 DiffNaps, an inherently interpretable binary neural network (binary autoencoder plus linear classification head over a binary pattern layer) that extracts class-specific differential patterns from high-dimensional binary data.
Problem
Finding succinct, non-redundant, interpretable patterns that differentiate classes in very high-dimensional data (e.g., gene expression), where existing pattern mining and rule-learning methods do not scale or are not robust to noise.
Approach
A binary autoencoder whose encoder/decoder share stochastically binarized weights so hidden neurons encode conjunctive patterns, combined with a linear classifier on the pattern layer; a multi-task loss (sparsity-aware reconstruction + classification) plus W-shaped and row-sum regularizers drives learning of patterns that are both descriptive and differential.
Key results
- On real data DiffNaps achieves the best AUC on Disease (.84), BRCA-S (.86) and Genomes (.77), and second best on Cardio (.62 for CART vs .56) and BRCA-N (.91 vs .95 for SPuManTe).
- DiffNaps scales to 225k features (Genomes) in about 9 minutes, while Classy takes 8h20m and SPuManTe/Premise fail on that dataset.
- On synthetic data DiffNaps stays robust up to 20% destructive noise, with F1 virtually unaffected, while competitors degrade.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DiffNaps instead of Cart | Class-specific pattern discovery · Cardio | AUC: .56 vs .62 | Hurt |
| Use DiffNaps instead of Premise | Class-specific pattern discovery · Disease | AUC: .84 vs .84 | No clear effect |
| Use DiffNaps instead of SPuManTe | Class-specific pattern discovery · BRCA-N | AUC: .91 vs .95 | Hurt |
| Use DiffNaps instead of Cart | Class-specific pattern discovery · BRCA-S | AUC: .31 -> .86 | Helped |
| Use DiffNaps instead of Cart | Class-specific pattern discovery · Genomes | AUC: .46 -> .77 | Helped |
| Use DiffNaps rather than competing methods as feature dimensionality increases | Synthetic pattern recovery · Synthetic data, varying feature count | F1 | Helped |
| Use DiffNaps rather than competing methods as destructive noise increases | Synthetic pattern recovery · Synthetic data, destructive noise up to 20% | F1: 20% | Helped |
Surprising findings
- None of the existing competing approaches are robust to destructive noise, according to the authors.
- CART and SPuManTe return thousands of patterns, undermining interpretability.
Limitations
- Only considers conjunctions of features as patterns, not more complex relations such as mutual exclusion.
- On low-dimensional data (up to ~100 features) DiffNaps performs only on par with other approaches.
- Ground truth is unknown for real data, so evaluation relies on proxy AUC and KEGG enrichment.
Experimental setup
Datasets: Cardio Disease BRCA-N (derived from TCGA) BRCA-S (derived from TCGA) Genomes synthetic data
Benchmarks: synthetic planted patterns Cardio Disease BRCA-N BRCA-S Genomes
Metrics: soft F1 AUC of sensitivity-specificity curve runtime number of patterns average pattern length
Baselines: CART Classy Premise SPuManTe RLL
Builds on: decision trees W-shaped regularizer straight-through estimator gated STE
Closest papers by meaning
- Codebook Features: Sparse and Discrete Interpretability for Neural Networks (ICML 2024 · similarity 0.82)
- Rank-Transformed Dissimilarity Profiles for High-Dimensional Classification (NeurIPS 2026 · similarity 0.81)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.81)
- MAGE: Model-Level Graph Neural Networks Explanations via Motif-based Graph Generation (ICLR 2025 · similarity 0.81)
- Concept-Guided Interpretability via Neural Chunking (NeurIPS 2025 · similarity 0.81)
- Derivative-Free Guidance in Continuous and Discrete Diffusion Models with Soft Value-based Decoding (NeurIPS 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.