PORTAL: Scalable Tabular Foundation Models via Content-Specific Tokenization

TL;DR

PORTAL is a scalable self-supervised framework for tabular data that uses content-specific tokenization, enabling effective pre-training on diverse, uncleaned datasets and achieving competitive results on classification and regression tasks.

Contribution

The paper introduces PORTAL, a novel content-specific tokenization method that allows scalable pre-training of tabular models without data cleaning or structural constraints.

Findings

Effective pre-training on large, uncleaned datasets

Achieves state-of-the-art results on complex tasks

Handles multiple data modalities without preprocessing

Abstract

Self-supervised learning on tabular data seeks to apply advances from natural language and image domains to the diverse domain of tables. However, current techniques often struggle with integrating multi-domain data and require data cleaning or specific structural requirements, limiting the scalability of pre-training datasets. We introduce PORTAL (Pretraining One-Row-at-a-Time for All tabLes), a framework that handles various data modalities without the need for cleaning or preprocessing. This simple yet powerful approach can be effectively pre-trained on online-collected datasets and fine-tuned to match state-of-the-art methods on complex classification and regression tasks. This work offers a practical advancement in self-supervised learning for large-scale tabular data.