Doctoral Dissertation Oral Proposal, Yushan Jiang

Title: Empowering Time Series Analysis with Knowledge Discovery and Inference
Ph.D. Candidate: Yushan Jiang
Major Advisor: Dr. Dongjin Song
Associate Advisors: Dr. Jinbo Bi, Dr. Derek Aguiar
Date/Time: Monday, May 13th, 2024, 10:00 am
Location: HBL1947 room
Meeting link: https://uconn-cmr.webex.com/uconn-cmr/j.php?MTID=m8a8d03614b0d9acd508ac42bd0c03037
Meeting number: 2631 460 5779
Password: aPXYCWFK693

Abstract:
Multivariate time series (MTS) analysis is crucial to unveil and comprehend temporal system behaviors and assist downstream decision-making with predictive insights. However, the complex dependencies and evolving natures of MTS present significant analytical challenges. MTS data can exhibit multi-mode behaviors over time, which requires additional analytic power to handle the variability with reliable explanations. Moreover, MTS data constantly evolves, leading to the disparity of underlying dependencies across different stages, which makes the standard analytic model less effective.
To cope with these challenges, we propose a knowledge discovery and utilization framework to empower the MTS analysis under the deep learning framework. MTS analysis can benefit from the integration of ubiquitous knowledge, either internal knowledge derived from MTS itself or external knowledge reflecting specific intra-domain and cross-domain insights. The first proposed approach discovers prototypical knowledge to tackle and interpret the variability in complex temporal sequences, which is deployed in an imitation learning task for dynamic treatment recommendation. Based on the prototype discovery of treatment trajectories, the model is able to provide case-based reasoning for varying patient symptoms and more accurate recommendations even with data heterogeneity. Beyond the holistic multivariate modeling of temporal knowledge, the second approach further explicitly leverages external structural knowledge representing variable interactions for a continual MTS forecasting task, where different regimes of MTS accumulate as the temporal system evolves or switches. By steering the model toward identifying the structural knowledge, and replaying effective samples based on the learned representations, the model is able to memorize the learned knowledge thus maintaining forecasting performance over all historical data.
Besides the basic view of MTS, we are currently working on knowledge discovery from other domains and other modalities that provide unique perspectives of temporal system behaviors, such as the frequency domain knowledge that complements the information based on Fourier components, and the cross-modal knowledge that provides foundational sequence modeling knowledge from large language models.