CRII: III: Learning Dynamic Graph-based Precursors for Event Modeling

点击翻译按钮获取中文摘要

From epidemic outbreaks to civil strife, societal events that involve large populations often deeply affect people’s lives and cause economic burden. Forecasting these events while providing context analysis helps social scientists and health practitioners to interpret and study human societies. Although many existing research efforts strive to forecast societal events, providing structured explanations for prediction is still limited given the underlying connections among entities, actions, and locations behind these events. This project presents a novel paradigm of identifying and organizing multiple types of precursors while predicting events. It identifies changing relations among entities as events evolve and studies the hidden geographical influence on events. Both entity relations and geographical connections are represented by dynamic graphs. Organizing event precursors in graphs greatly reduces the complexity of comprehending unstructured input data and delivers interpretable summarizations for event prediction. This work will involve educational activities such as development of course curriculum; training of graduate, undergraduate, and high-school students; encouraging participation of women and minority groups in academic research; and dissemination of outcomes such as software and datasets for the general public.To achieve these goals, this project will integrate multiple data sources and analyze complex hierarchical features in modeling events. Although a variety of online data has been utilized to analyze and predict societal events, it also raises new challenges such as: (1) accounting for dynamic relationships within data sets; (2) preserving and learning complex knowledge structures with heterogeneous data sets; and (3) ensuring interpretable results for predictions and decision making. This project will address the challenges in the following ways: (i) it will integrate multi-source data by learning a unified multi-level semantic encoding; (ii) it will identify historical key semantics by paying attention to hierarchical text structures in a recurrent learning process; (iii) it will provide explanations for event prediction by incorporating local dynamic graph patterns and global influence graph patterns. The specific research aims will be complemented with an extensive set of evaluation plans including a retrospective evaluation on real-word event records and a user survey to evaluate graph visualizations of event precursors. The project results, including graph based empirical data, predictive evaluation tools, and open source software for analyzing events, will be shared with computer science research community and stakeholders in computational healthcare, and social science.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.