IPSI: Enhancing Structural Inference with Automatically Learned Structural Priors

Zhongben Gong, Xiaoqun Wu, Mingyang Zhou

Keywords: Structural Inference, AI for Science, Graph Neural Networks

TL;DR: We propose IPSI, a general structural inference framework that improves performance by injecting pretrained structural priors into variational inference module.

Abstract:

We propose IPSI, a general iterative framework for structural inference in interacting dynamical systems. It integrates a pretrained structural estimator and a joint inference module based on the Variational Autoencoder (VAE); these components
are alternately updated to progressively refine the inferred structures. Initially, the
structural estimator is trained on labels from either a meta-dataset or a baseline
model to extract features and generate structural priors, which provide multi-level
guidance for training the joint inference module. In subsequent iterations, pseudolabels from the joint module replace the initial labels. IPSI is compatible with
various VAE-based models. Experiments on synthetic datasets of physical systems
demonstrate that IPSI significantly enhances the performance of structural inference models such as Neural Relational Inference (NRI). Ablation studies reveal
that feature and structural prior inputs to the joint module offer complementary
improvements from representational and generative perspectives.

Primary Area: Machine learning for sciences (e.g. climate, health, life sciences, physics, social sciences)

Submission Number: 15510