Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Marked Neural Spatio-Temporal Point Process Involving a Dynamic Graph Neural Network

Alice Moallemy-Oureh, Silvia Beddar-Wiesing, Yannick Nagel, Rüdiger Nather, Josephine M. Thomas

TL;DR

The paper tackles learning on fully dynamic graphs where events alter structure and attributes over time. It introduces the Marked Neural Spatio-Temporal Point Process (MNSTPP), which combines a Dynamic Graph Neural Network with a separate neural TPP for each event type, enabling prediction of event times, types, and attribute changes. Key contributions include the first integration of a DGNN with marked neural TPPs for graph streams, per-event-type scoring and decay mechanisms, online update capability via local retraining, and a comprehensive training/prediction framework with negative sampling and batch processing. This framework broadens the applicability of temporal graph learning to diverse scientific domains where graphs evolve continually and attributes change, facilitating real-time forecasting and analysis of complex dynamic systems.

Abstract

Temporal Point Processes (TPPs) have recently become increasingly interesting for learning dynamics in graph data. A reason for this is that learning on dynamic graph data is becoming more relevant, since data from many scientific fields, ranging from mathematics, biology, social sciences, and physics to computer science, is naturally related and inherently dynamic. In addition, TPPs provide a meaningful characterization of event streams and a prediction mechanism for future events. Therefore, (semi-)parameterized Neural TPPs have been introduced whose characterization can be (partially) learned and, thus, enable the representation of more complex phenomena. However, the research on modeling dynamic graphs with TPPs is relatively young, and only a few models for node attribute changes or evolving edges have been proposed yet. To allow for learning on fully dynamic graph streams, i.e., graphs that can change in their structure (addition/deletion of nodes/edge) and in their node/edge attributes, we propose a Marked Neural Spatio-Temporal Point Process (MNSTPP). It leverages a Dynamic Graph Neural Network to learn a Marked TPP that handles attributes and spatial data to model and predict any event in a graph stream.

Marked Neural Spatio-Temporal Point Process Involving a Dynamic Graph Neural Network

TL;DR

The paper tackles learning on fully dynamic graphs where events alter structure and attributes over time. It introduces the Marked Neural Spatio-Temporal Point Process (MNSTPP), which combines a Dynamic Graph Neural Network with a separate neural TPP for each event type, enabling prediction of event times, types, and attribute changes. Key contributions include the first integration of a DGNN with marked neural TPPs for graph streams, per-event-type scoring and decay mechanisms, online update capability via local retraining, and a comprehensive training/prediction framework with negative sampling and batch processing. This framework broadens the applicability of temporal graph learning to diverse scientific domains where graphs evolve continually and attributes change, facilitating real-time forecasting and analysis of complex dynamic systems.

Abstract

Temporal Point Processes (TPPs) have recently become increasingly interesting for learning dynamics in graph data. A reason for this is that learning on dynamic graph data is becoming more relevant, since data from many scientific fields, ranging from mathematics, biology, social sciences, and physics to computer science, is naturally related and inherently dynamic. In addition, TPPs provide a meaningful characterization of event streams and a prediction mechanism for future events. Therefore, (semi-)parameterized Neural TPPs have been introduced whose characterization can be (partially) learned and, thus, enable the representation of more complex phenomena. However, the research on modeling dynamic graphs with TPPs is relatively young, and only a few models for node attribute changes or evolving edges have been proposed yet. To allow for learning on fully dynamic graph streams, i.e., graphs that can change in their structure (addition/deletion of nodes/edge) and in their node/edge attributes, we propose a Marked Neural Spatio-Temporal Point Process (MNSTPP). It leverages a Dynamic Graph Neural Network to learn a Marked TPP that handles attributes and spatial data to model and predict any event in a graph stream.
Paper Structure (24 sections, 20 equations, 1 figure, 1 table)

This paper contains 24 sections, 20 equations, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. Foundations
  3. Graph Streams
  4. Temporal Point Processes
  5. Marked Temporal Point Processes.
  6. Neural Temporal Point Processes.
  7. Model
  8. Dynamic Graph Neural Network
  9. Scoring Functions
  10. Node Addition.
  11. Node Deletion.
  12. Edge Addition.
  13. Edge Deletion.
  14. Node Attribute Change.
  15. Edge Attribute Change.
  16. ...and 9 more sections

Figures (1)

  • Figure 1: The figure is inspired by STPP_book. The shaded rectangle is the ball for which we want to have the probability and the dots over time visualize the observed events.

Theorems & Definitions (3)

  • Definition 2.1: Marked Temporal Point Process
  • Remark 2.2: Spatio-Temporal Point Process
  • Definition 2.3: Neural Temporal Point Process