Amortized Causal Discovery with Prior-Fitted Networks

TL;DR

This paper introduces a novel amortized causal discovery method using Prior-Fitted Networks to improve likelihood estimation accuracy, leading to better causal structure recovery across various datasets.

Contribution

The paper presents a new approach leveraging PFNs for amortized likelihood estimation, addressing errors in traditional methods and enhancing causal discovery accuracy.

Findings

PFNs provide more accurate likelihood estimates than conventional neural networks.

The method significantly improves structure recovery on synthetic and real datasets.

Experiments demonstrate better performance over standard baselines.

Abstract

In recent years, differentiable penalized likelihood methods have gained popularity, optimizing the causal structure by maximizing its likelihood with respect to the data. However, recent research has shown that errors in likelihood estimation, even on relatively large sample sizes, disallow the discovery of proper structures. We propose a new approach to amortized causal discovery that addresses the limitations of likelihood estimator accuracy. Our method leverages Prior-Fitted Networks (PFNs) to amortize data-dependent likelihood estimation, yielding more reliable scores for structure learning. Experiments on synthetic, simulated, and real-world datasets show significant gains in structure recovery compared to standard baselines. Furthermore, we demonstrate directly that PFNs provide more accurate likelihood estimates than conventional neural network-based approaches.