CRII: III: Interpretable Models for Spatio-Temporal Event Forecasting using Social Sensors

点击翻译按钮获取中文摘要

Significant events that occur at certain times and in specific locations, such as disease out-breaks and crime incidents, have tremendous impacts on our society. This strongly motivates the need to anticipate event occurrences in advance in order to reduce the potential social upheaval and damage caused. For example, for traffic congestion predictions based on social network reporting and traffic sensors, these methods would inform the authorities where the future congestion will occur and why certain ongoing traffic incident and congestion hot spots will worsen the problem on specific roads. In recent years, such model interpretability has attracted increasing attention as machine learning is beginning to be applied to ever more practical applications. As a domain with significant impact on society, the interpretability of spatio-temporal event forecasting models is particularly important in order to earn the trust of practitioners and become widely adopted in their everyday workflow. However, like conventional machine learning models, models for social event forecasting still primarily focus on prediction accuracy and are rapidly becoming too sophisticated and obscure to be easily understood by human operators. There is thus an urgent need to fill the increasing gap between data scientists and practitioners. To address it, this project focuses on developing a novel spatio-temporal social event forecasting framework that can jointly optimize the model accuracy and interpretability, and automatically illustrate the explanatory process of prediction generation. To address challenges like spatial dependency and high-dimensional large data, the project aims at exploring the conditional independence and spatial topology to boost the sparsity of spatial dependence patterns. The project will then move on to exploit the hierarchical conjunction lattice of primitive data features to enforce the conciseness and sparsity of expository high-level representations of the data. To solve the formulated optimization problem for jointly maximizing accuracy and interpretability, this project also involves research on the corresponding optimization methods with rigorous theoretical analysis on efficiency and optimality. Finally, strategies for evaluating model interpretability in social event forecasting are systematically investigated. The success of this project will shed a light on the generic research in interpretable data mining and machine learning. The methods and tools developed in this project will help fill the gaps between data scientists and domain-specific forecasting experts. Finally, this project will provide valuable resources to support courses with new topics, datasets, techniques, and software, and gives more research opportunities for underrepresented students.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.