Workshop Overview

This virtual workshop was meant to attract participants with a research interest in explainable AI within the Geographic domain (GeoAI). The event focused on exploring the reasoning path of GeoAI in the analysis of geographic phenomena, finding ways to bring existing spatial theories and models closer to the rapidly evolving data-intensive science, and identifying directions for future research that will promote explainability in GeoAI.

Topic & Goals

Recent advances in artificial intelligence (AI) research, the significant increase in computational power, and the large-scale availability of data have ushered in a new era of data-intensive science. In the context of GIScience, GeoAI aims to employ AI methods to analyze complex geographical phenomena. Examples include the detection of terrain features and predictive analytics on spatial patterns of various phenomena in crime activity, economy, health, and human behavior. The majority of GeoAI applications rely on machine learning (ML) models to extract generalizable predictive patterns, in the form of mathematical models, that provide useful insights about the phenomenon in question. Machine learning excels in efficiency, scalability, and accuracy; however, this comes at the cost of reduced explainability. A clear reasoning path from data to conclusions is not always evident, but is readily available in the traditional analysis of geographic phenomena using conceptual and statistical models.

This workshop aimed to bring together GeoAI researchers/data analysts with experts in explainability as well as experts from the domains of human mobility, social sciences, environmental psychology, and criminology. On one hand, GeoAI researchers/data analysts had the opportunity to present and evaluate their findings by taking into consideration the knowledge of domain experts based on formal theories and models. On the other hand, domain experts (criminologists, environmental psychologists, and so forth) had the opportunity to learn about capabilities provided by GeoAI and data science that may challenge the classical patterns of explanation that exist in formal models. Indicatively, environmental psychology experts were able to learn about state-of-the-art sentiment analysis techniques to discover hotspots of positive/negative emotional responses. At the same time, researchers that use these techniques can evaluate how well they explain human behavior by leveraging the expertise of environmental psychologists.

Structure and outcome

The workshop took place during the week of October 12, 2020, organized as three 1-hour Zoom sessions, with the objective of providing opportunities for networking and further collaboration among researchers from diverse backgrounds.

Following the workshop, participants will be invited to submit their work to a special issue published in Transactions in GIS.

Discussion Panel

Name AffiliationArea
Ben AdamsUniversity of CanterburyArtificial Intelligence
Bruno MartinsUniversity of LisbonArtificial Intelligence
Song GaoUniversity of Wisconsin-MadisonArtificial Intelligence
Constantinos AntoniouTechnical University of MunichHuman mobility
Nick MallesonUniversity of LeedsCrime Behavior
Mehul BhattÖrebro UniversityAI, Spatial Cognition, Reasoning

Timetable

TitleAuthors

October 15th @ 12:00 - 1:00 p.m. (PT)
Crime Activity - Panel: Nick Malleson, Bruno Martins

Machine Learning or Bust: Does The Criminological Community Need to Adopt More Advanced Techniques For Crime Forecasting?Justin Kurland

October 16th @ 12:00 - 1:00 p.m. (PT)
Travel Behavior - Panel: Constantinos Antoniou, Song Gao

A Traffic Flow Forecasting for Scenic Spots based on Multi-source and Heterogeneous DataYuan Gao , Yuan-Lei Shi and Yao-Yi Chiang
Graph Markov network for traffic forecasting with missing dataZhiyong Cui
Deep learning for travel behaviour: When and how?Melvin Wong and Bilal Farooq

October 20th @ 4:00 - 5:00 p.m. (PT)
Invited Talk - Ben Adams: Contrastive explanations in GeoAI

October 21st @ 12:00 p.m. - 1:00 p.m. (PT)
Society & Env. Psychology - Panel: Mehul Bhatt, Bruno Martins

A Random Forest Based Spatiotemporal Regression Model for the Impact of Green Space on Childhood Academic Performance
Bita Minaravesh and Orhun Aydin
GeoAI for Spatially Explicit Population ProjectionsMarina Georati and Carsten Kessler
Rotational Locomotion in Built-up Spaces. Implications for Evidence-Based DesignVasiliki Kondyli and Mehul Bhatt

Attendance

The workshop was organized as three separated sessions: Crime Activity, Travel Behavior, and Society/Env. Psychology.

  • Human Mobility
    • Estimating travel behavior
    • Traffic forecasting
  • Crime Behavior
    • Perceived safety in cities
    • Hate crime and disorder
  • Social Sensing
    • Place recommendation systems
    • Human activities and emotions
  •