Imputing Incomplete Building Attribute Data Using Expert-Guided Variational Autoencoders (EGVAE)

ICLR 2026 Conference Submission21751 Authors

Keywords: Building Attribution, Data Imputation, Variational AutoEncoder, Causal Graph, Expert Priors, Neural Networks

TL;DR: We propose a building attribute imputation method when no or limited training data is available. In such scenarios existing ML algorithms fail to generate imputations, our method leverages expert opinions to generate sensible imputations

Abstract: Human activity is organized by the physical infrastructure such as roads and buildings, making their characterization essential for applications ranging from disaster preparedness and national security to real estate analytics and public resource allocation. Despite the availability of datasets detailing building attributes many of these are incomplete, particularly in data scarce regions, limiting their utility in critical decision making tasks. We propose a deep learning approach for imputing missing building attributes by learning from sparse to no observed data, expert knowledge, and spatial correlations among buildings. Our model is based on a Vector Quantised-Variational AutoEncoder (VQ-VAE) architecture with a graph neural network (GNN) encoder that captures spatial dependencies, while an additional KL-divergence based loss term incorporates expert-informed priors. By jointly leveraging observed data and expert-informed priors, the model learns latent representations that enable imputing missing data for attributes with little or no training data. Experimental results on real-world datasets demonstrate the robustness and effectiveness of our proposed method.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 21751