Research map / Label-efficient and robust learning / Recognition and long-tail
Interpretable Deep Clustering 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 IDC, a two-stage deep clustering method for tabular data that learns sample-specific sparse feature gates via a self-supervised gated autoencoder and cluster-level gates via an auxiliary classifier, yielding instance- and cluster-level interpretations.
Problem
Clustering tabular data (especially high-dimensional biomedical data) is hard, and existing deep clustering methods rely on image augmentations and offer no sample-level or cluster-level feature interpretability.
Approach
Learn sample-specific stochastic gates (local feature selection) through a denoising autoencoder trained with a gated reconstruction loss and a gates total coding rate term, then cluster the sparsified embeddings with a coding-rate clustering head and derive cluster-level gate matrix from an auxiliary classifier.
Key results
- On 11 real tabular datasets, IDC achieves best or second-best clustering accuracy, best on 9 of 11 (e.g., ALLAML 77.5 vs 74.5 for next best DUFS).
- On MNIST-10K, IDC with gates reaches 83.45% ACC with faithfulness 0.96 and generalizability 88.5, outperforming SHAP+K-means (53.34% ACC).
- On MNIST-60K, IDC reaches 87.9% ACC using about 16 selected features out of 784.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use IDC instead of K-means + SHAP | Synthetic clustering · Synthetic | ACC: 25.72\% -> 99.91\% | Helped |
| Add local gates (ablation) | IDC · MNIST60K | ACC: 81.1 -> 87.9 | Helped |
| Add local gates (ablation) | IDC · FashionMNIST | ACC: 61.0 -> 61.9 | Helped |
| Add gate regularization L_reg (ablation) | IDC · MNIST60K | ACC: 85.9 -> 87.9 | Helped |
| Add latent denoising (ablation) | IDC · MNIST60K | ACC: 86.5 -> 87.9 | Helped |
| Add input denoising (ablation) | IDC · MNIST60K | ACC: 84.3 -> 87.9 | Helped |
| Use the learned clustering head instead of K-means (ablation) | IDC features · MNIST60K | ACC: 65.5 -> 87.9 | Helped |
| Add reconstruction loss L_recon (ablation) | IDC · MNIST60K | ACC: 18.0 -> 87.9 | Helped |
| Use IDC instead of TELL | CIFAR10 clustering · CIFAR10 | ARI: 5.96 -> 6.16 | Helped |
| Use IDC instead of TELL | CIFAR10 clustering · CIFAR10 | ACC: 25.65 -> 25.01 | Hurt |
Surprising findings
- Removing the gates' regularizer (lambda_reg=0) gave higher accuracy on the MNIST-10K sensitivity study (86.4%) while keeping far more open gates (435).
Limitations
- Struggles with correlated variables, a known caveat of reconstruction loss.
- Does not handle datasets with a large number of clusters well.
Experimental setup
Datasets: MNIST FashionMNIST CIFAR-10 TOX-171 ALLAML PROSTATE SRBCT BIASE INTESTINE PBMC-2 CNAE-9 MFEATZ
Benchmarks: MNIST-10K interpretability TOX-171 ALLAML PROSTATE SRBCT BIASE INTESTINE PBMC-2 CNAE-9 MFEATZ MiniBooNE ALBERT
Metrics: ACC ARI NMI uniqueness diversity faithfulness generalizability number of selected features
Baselines: K-means TELL k-DVAE SRCFS CAE DUFS LS MCFS VaDE DEC SHAP Integrated Gradients
Builds on: stochastic gates maximum coding rate reduction TELL denoising autoencoder Gumbel-Softmax
Closest papers by meaning
- ZEUS: Zero-shot Embeddings for Unsupervised Separation of Tabular Data (NeurIPS 2025 · similarity 0.85)
- Interpretable Mesomorphic Networks for Tabular Data (NeurIPS 2024 · similarity 0.85)
- Hybrid Autoencoders for Tabular Data: Leveraging Model-Based Augmentation in Low-Label Settings (NeurIPS 2025 · similarity 0.84)
- TabGLM: Tabular Graph Language Model for Learning Transferable Representations Through Multi-Modal Consistency Minimization (AAAI 2025 · similarity 0.84)
- TabSTAR: A Tabular Foundation Model for Tabular Data with Text Fields (NeurIPS 2025 · similarity 0.84)
- From Logits to Hierarchies: Hierarchical Clustering made Simple (ICML 2025 · 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.