Measuring disentangled generative spatio-temporal representation

TL;DR

This paper evaluates disentangled representation learning methods on spatio-temporal data, demonstrating their ability to maintain prediction performance while revealing meaningful spatial-temporal semantics.

Contribution

It introduces an internal evaluation metric for disentanglement in spatio-temporal models and applies state-of-the-art methods to large datasets, highlighting their interpretability.

Findings

Disentangled representations achieve comparable forecasting accuracy to existing methods.

The proposed metric effectively measures correlation among latent variables.

Methods reveal meaningful spatial-temporal semantics in learned features.

Abstract

Disentangled representation learning offers useful properties such as dimension reduction and interpretability, which are essential to modern deep learning approaches. Although deep learning techniques have been widely applied to spatio-temporal data mining, there has been little attention to further disentangle the latent features and understanding their contribution to the model performance, particularly their mutual information and correlation across features. In this study, we adopt two state-of-the-art disentangled representation learning methods and apply them to three large-scale public spatio-temporal datasets. To evaluate their performance, we propose an internal evaluation metric focusing on the degree of correlations among latent variables of the learned representations and the prediction performance of the downstream tasks. Empirical results show that our modified method can learn disentangled representations that achieve the same level of performance as existing state-of-the-art ST deep learning methods in a spatio-temporal sequence forecasting problem. Additionally, we find that our methods can be used to discover real-world spatial-temporal semantics to describe the variables in the learned representation.