Tabular clustering drops per-dataset tuning: TabClustPFN returns both the partition and the cluster count in one forward pass
Synopsis
TabClustPFN turns tabular clustering into a single forward pass: a pretrained network outputs each sample's cluster assignment and the number of clusters at once, without fitting a model per dataset. Earlier prior-fitted-network clustering methods either required the cluster count as input or avoided label switching by ordering clusters by distance to the origin, and their pretraining priors covered only low-dimensional Gaussian mixtures. On 44 real tabular datasets, the model ranks first in median ARI when the cluster count is known and second when it infers the count itself, with a median absolute error of about 1 in the inferred count.
Interpretation
Tabular clustering can now produce both the cluster assignment and the cluster count in one forward pass, with no per-dataset optimization or tuning. Earlier prior-fitted-network clustering methods addressed only part of the unsupervised difficulty: ZEUS learns tabular embeddings but hands clustering to k-means with a known cluster count, while Cluster-PFN removes label switching by hard-ordering clusters by their distance to the origin. On 44 real tabular datasets, TabClustPFN with the known cluster count achieves the best median ARI rank, and the variant that infers the count through its own network ranks second; the inferred count has a median absolute error of about 1, lower than other prior-fitted-network methods.
The model splits cluster assignment and cluster count into two networks, the first inferring a partition for a given count and the second reading the count back from partition statistics. The partition inference network keeps K learnable cluster prototypes that are iteratively co-refined with data representations through cross-attention, then turns normalized cosine similarity into soft assignments with a softmax; the cardinality inference network summarizes co-assignment statistics across candidate counts and passes them to a multilayer perceptron. Pretraining data mixes controlled-overlap Gaussian mixtures with clusters distorted by invertible residual-network transformations, covering mixed numerical and categorical features; the training objective is a differentiable relaxation of the Adjusted Rand Index that is permutation-invariant by construction.
Perspective
The demonstrated scope is tabular datasets with 2 to 10 clusters, at most 64 features and at most 1000 samples, with the real benchmark comprising 44 datasets filtered from sources including OpenML and UCI; sample sizes and feature dimensions beyond that range are covered by a separate out-of-distribution evaluation of the frozen checkpoint.
Direct extrapolation to larger feature dimensions is harder, mainly because cardinality estimation degrades after dimensionality reduction, which lowers ARI in the unknown-count setting; on high-difficulty synthetic datasets every method's performance drops markedly, and this model is no exception.
