Grant: Discovery Grants Program - Individual

This program aims to investigate the modelling of co-evolving sequential data and develop novel learning methods for predicting long-range future values or events. The research is central to our applications. In the healthcare domain, we study health trajectories involving repeated measures of risk factors to follow particular subjects over a prolonged period. Our aim is to predict clinical failure events such as death or rehospitalization. In the finance domain, we analyze time series comprising longitudinal trajectories of daily historical stock quotes, bond returns etc. Our aims are to identify regimes within the time series and predict regime changes to aid investment decision making. In social network analysis, we study the evolution of dynamic online networks. Our aims are to identify complex relationships between users, such as mutual influence and hidden affinities; to predict undesirable behaviours; and to detect changes in community structure. The long-term goal of this program is to build and validate frameworks to effectively mine spatiotemporal sequential patterns and generate predictive inferences while contributing to the development of more interpretable AI. Our short-term objectives are (1) to better represent co-evolving sequence data by meaningful patterns, profiles and trajectories; (2) to model interactions between co-evolving sequences in an ecosystem to improve regime detection, regime change prediction and sequence/event prediction; (3) to investigate outlier dynamics to better understand the driving forces for regime changes; (4) to develop effective survival learning algorithms for learning from trajectories and predicting long-range events; and (5) to develop methods for elucidating the decision process. This program will be carried out by accomplishing six interrelated projects. Project 1 focuses on discovering rich patterns that can be used as the building blocks for representing co-evolving trajectories. Projects 2 and 3 aim to develop flexible models of interactions via community analysis approaches. For this purpose, we propose to build a time-evolving network graph from segments or windows of trajectories, and investigate the evolving community structures within the network. Project 2 is about mining customers' electricity consumption behaviour for load forecasting, while Project 3 is about mining regime events for time series forecasting. Project 4 is about investigating outlier dynamics causing structural breaks in the evolution of individual trajectories or groups of trajectories. Project 5 attempts to develop survival-learning machines that explore the underlying relationship between repeated measures of covariates and failure-free survival probability. Finally, Project 6 investigates survival analysis from the perspective of a heterogeneous information network of longitudinal data in order to develop healthcare trajectory applications. This program will train 10 HQPs.