Explainable Machine Learning for Activity Modeling in GeoAI
Date
2020
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
GeoAI is a recent cutting-edge discipline that combines advancements in Big Geospatial Data Management (Geo) and Artificial Intelligence (AI). This thesis focuses on two GeoAI aspects, (1) AI for Geo, which applies advanced machine-learning-based models to emerging geospatial data; and (2) Geo for AI, which proposes novel machine learning algorithms that have better explainability aligned with geospatial knowledge. AI for Geo applies machine learning to a wealth of emerging spatiotemporal datasets generated by users, such as OpenstreetMap, Twitter, Yelp, or farecard data from transportation systems. This data captures user activities and the dynamics of a changing environment. Our approach to derive knowledge is based on so-called Activity Modeling for spatiotemporal data by proposing specific machine learning methods to tackle the following novel challenges.
(i) The data quality and validity of user-generated data are still of some concern and we show the feasibility of using Openstreetmap Edits for assessing urban change using statistical modeling.
(ii) Another challenge is using explainable latent temporal patterns in machine learning models, e.g., the spatial proximity and temporal auto-correlation for user clustering and trajectory synthesis;
(iii) The last challenge is how to introduce new data representation (such as continuous-time temporal graphs) and latest deep learning methods (including Factorized Variational Autoencoder and Generative Adversarial Neural Networks) to capture complex high-dimensional information beyond conventional data representation and methods. From the perspective of Geo for AI, machine learning (ML) models help solve many challenging problems such as computer vision, speech processing, and also spatiotemporal data. However, people are expecting good explainability of results for decision making (e.g. healthcare, law enforcement, and self-driving systems) or first-principle scientific domain knowledge (e.g. chemical bonds, physics movement, and biological linkage). As such, this thesis is also motivated by promoting machine learning explainability in Activity Modeling to generate and enforce better explainability for spatiotemporal data and specific applications beyond what is possible with generic models. Specifically, this work addresses
(i) the predictive modeling of urban change using a novel autoregression approach based on power-law growth principles and spatiotemporal auto-correlation,
(ii) explainable user clustering based on matrix factorization as part of a transfer learning framework leveraging tidal traffic and commuting behaviors,
(iii) spatiotemporal trajectory generation using Variantional Autoencoders through a factorized generative model and spatiotemporal-validity constraints on driving behavior and the physical limitations of vehicular movement,
and finally (iv) Generative Adversarial Network (GAN) based temporal graph generation using novel deep structures on temporal dynamics and location representation. Overall, this thesis comprises six chapters. Chapter 1 discusses GeoAI and Activity Modeling and highlights how the two perspectives of GeoAI motivate explainability in machine learning. Chapters 2 through 5 cover four specific issues related to Activity Modeling. Chapter 6 summarizes this thesis and identifies future works.