Spatio-temporal Similarity Relation Representation Learning based on Graph Neural Network

Abstract

With the advance of the Internet of Things (IoT), massive Spatio-Temporal (ST) data are collected, providing an unprecedented opportunity to study human mobility. Extracting similarity information from such vast ST data is crucial for data mining tasks, such as clustering, classification, prediction, etc. In modern Location Based Services (LBS), deep representation learning is employed to embed ST data into a low-dimensional vector space for extracting similarity information. Recent deep learning approaches leverage sequential models for similarity relation representation; however, these methods are inefficient and their performance is impacted by data length. In this work, we propose Graph-based Spatio-Temporal Representation (GSTR) which exploits similarity relation representation learning on the K-Nearest Neighbor (KNN) graph with Graph Neural Networks (GNNs), efficiently capturing the original ST similarity. The experiments demonstrate that GSTR outperforms the state-of-the-art baselines, including matrix factorization approaches and deep learning methods in terms of similarity preservation, dimension reduction and representation efficiency.