TabClustPFN opens one-pass clustering for unlabeled tables
Researchers from Renmin University, Rutgers University, the National University of Singapore and MBZUAI introduced TabClustPFN, an open-source pretrained network for clustering unlabeled tables. The model uses a single forward pass to predict sample assignments and candidate cluster counts, with probabilities indicating its confidence in each count; its code and pretrained weights are available. The paper reports strong ARI rankings on 44 real datasets and 96.8% coverage of the true cluster count by candidate sets built to 95% cumulative probability on 3,200 held-out synthetic datasets.
TabClustPFN combines automatic clustering, cluster-count uncertainty and one-pass inference in an open-source tabular model, reducing the iterative algorithm-selection and tuning workflow described by the research team.
机器之心 2026-10-25 13:28 北京
聚类还要反复调参吗？
拿到一张没有标签的表格，能不能直接找出其中的群体，还知道模型对 “分几组” 有多大把握？TabClustPFN 尝试用预训练后的单次前向推断，缩短从数据到分组结果的距离。
给你一张客户表，你能看出里面有几种不同的消费人群吗？有人买得频繁，但每次花得不多；有人平时不来，一来就是大单。表格记录了这些行为，却没有告诉你：哪些人应该归为一组，一共又该分成几组。
#TabClustPFN#Renmin University#Rutgers University#National University of Singapore#MBZUAI