Deep Learning Approaches for Modeling Spatio-Temporal Dynamics in Evolving Networks

Finding appropriate representations for spatio-temporal data using deep learning techniques has attracted considerable interest in the analysis of complex evolving networks, as the learned representations have demonstratedstate-of-the-art performance in addressing a variety of difficult tasks across a broad range of domains. The objective of this dissertation is to apply a data-driven learning strategy to analyze multi-dimensional time-course data in order to more accurately model time-varying networks, such as the intelligent transportation network and the communication infrastructure network. We develop unified models to examine the following problems: estimation of spatio-temporal missing data in the intelligent transportation system, representation learning on discrete time spatio-temporal graphs, which is further used for the missing data estimation task, and representation learning over discrete time temporal graphs using temporal point processes and generalized temporal Hawkes processes. In Chapter 3, we tackle the problem of missing data in spatio-temporal measurements in intelligent transportation systems (ITS), where portions of collected traffic speed and travel time estimations in ITS are missing due to sensor instability and communication errors at collection points. These practical issues can be remedied by missing data analysis, which is mainly categorized as either statistical or machine learning (ML)-based approaches. We focus on an ML-based approach, Multi-Directional Recurrent Neural Network (M-RNN). M-RNN utilizes both temporal and spatial characteristics of the data. We evaluate the effectiveness of this approach on a TomTom dataset containing spatio-temporal measurements of average vehicle speed and travel time in the Greater Toronto Area (GTA). In Chapter 4, we study the problem of representation learning on discrete time dynamic graphs, which are sequences of snapshots sampled from a dynamic graph at regularly-spaced times. We propose a Temporal Multilayer Position-Aware Graph Neural Network (TMP-GNN), a node embedding approach that incorporates the interdependence of temporal relations into embedding computation. Each layer in our model is a graph built from existing nodes and weighted edges corresponding to a given time. We learn the short-term temporal dependencies, global position, and feature information of the graph jointly through our TMP-GNN embedding component and utilize the derived representation in a missing data estimation framework. Additionally, we deploy the concept of conditional centrality derived from eigenvector-based centrality to distinguish nodes of higher influence and integrate it in message aggregation across the graph. We conduct several experiments using four real-world datasets to evaluate the performance of TMP-GNN on two different representations of temporal multilayered graphs. In Chapter 5 of this dissertation, we work on the critical problem of temporal graph representation learning, in which we acquire representations that change over time on a graph. We consider general structural changes in the graph, such as the formation or removal of a graph node or edge at a specific time, to be an event, and a graph evolves continuously as more events are added. On the other hand, such sequences of discrete events occur at irregular time scales and are thus modeled using a stochastic process, particularly temporal point processes (TPP). Our focus is to learn the conditional intensity function of the temporal point process in a data-driven manner to investigate the influence of deletion event types on representation learning of the nodes towards a link-level prediction task. In this regard, we extract local and graph-level measures, particularly network entropy, which quantifies the node/edge significance within the network, to capture the impact of node deletion (and corresponding edge deletion) and incorporate them into our integrated framework. Then, we study the correlation between two temporal point processes, each modeling addition types of events (network growth) and deletion types of events (network shrinkage), and observe the statistically significant asynchrony (repulsive) behavior of the processes towards each other. Following that, we work on the generalized temporal Hawkes processes and develop an adaptive representation learning algorithm to model the dependency between the network growth and shrinkage events.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.