TabClustPFN: Out-of-the-Box Tabular Clustering with Probabilistic Cluster Count

Traditional clustering methods like K-means require iterative optimization and manual hyperparameter tuning for each new dataset, including the number of clusters. This creates a barrier for rapid exploratory analysis.

TabClustPFN Approach

TabClustPFN (Tabular Clustering Prior-data Fitted Network) shifts the learning burden to a pre-training phase. The model is trained on a large collection of synthetic tabular tasks generated from two priors: Gaussian Mixture Models (controllable cluster shapes and overlaps) and ZEUS (nonlinear transformations covering numerical and categorical features). Because the generative process is known, the true cluster assignments are available for supervision.

Architecture

The network comprises two modules:

Partition Inference Network (PIN) : predicts sample assignments for each candidate cluster count (2 to 10).

Cluster-count Inference Network (CIN) : aggregates structural information from PIN's partitions to output a probability distribution over the candidate cluster counts.

Training uses a differentiable SoftARI objective that measures partition quality independent of cluster label permutations, avoiding the label-switching problem inherent in clustering.

Inference

At test time, the model parameters are frozen. A new table passes through the network once (forward pass) to produce both cluster assignments and a probability distribution over the number of clusters. This eliminates per-dataset training and iterative optimization.

Empirical Evaluation

Speed vs. Quality

On benchmark tasks, TabClustPFN's runtime is comparable to simple baselines and faster than other PFN-based clustering methods, while maintaining strong clustering performance (measured by Adjusted Rand Index).

Real-World Datasets

Evaluated on 44 real datasets (22 from OpenML, 22 from UCI) with up to 1,000 samples, 64 features, and 10 classes:

When given the true cluster count, TabClustPFN ranks first in median ARI.

When estimating the cluster count automatically, it ties for second in median ARI, showing minimal degradation.

Cluster-count estimation median absolute error is 1, outperforming other PFN-based approaches.

Uncertainty Calibration

The model outputs a full probability distribution over cluster counts (2–10). On 3,200 held-out Gaussian mixture datasets, the 95% cumulative probability set covers the true cluster count 96.8% of the time, indicating well-calibrated uncertainty. Examples show sharp distributions for well-separated clusters and diffuse distributions for ambiguous structures.

Availability

Paper: TabClustPFN: A Prior-data Fitted Network for Tabular Data Clustering (NeurIPS 2026). Authors: Tianqi Zhao (first author, Renmin University), Guanyang Wang, Yan Shuo Tan, Qiong Zhang (corresponding author, MBZUAI). Code and pre-trained weights: https://github.com/Tianqi-Zhao/TabClustPFN arXiv: https://arxiv.org/abs/2601.21656