Long-Tailed Learning in the Open and Dynamic World: Theories, Algorithms, and Applications

Presenter: Dawei Zhou

Bio: Dawei Zhou is an Assistant Professor at the Computer Science Department of Virginia Tech and the director of the Virginia Tech Learning on Graphs (VLOG) Lab. Zhou’s primary research focuses on open-world machine learning, with applications in hypothesis generation and validation, financial fraud detection, cyber security, risk management, predictive maintenance, and healthcare. He obtained his Ph.D. degree from the Computer Science Department of the University of Illinois Urbana-Champaign (UIUC). He has authored more than 60 publications in premier academic venues across AI, data mining, and information retrieval (e.g., ICML, NeurIPS, AAAI, IJCAI, KDD, ICDM, SDM, TKDD, DMKD, WWW, CIKM) and has served as Vice Program Chair/Proceeding Chair/Local Chair/Social Media and Publicity Chair/Session Chairs/(Senior) Program Committee Members in various top ML and AI conferences (e.g., KDD, NeurIPS, ICML, WWW, SIGIR, ICLR, AAAI, IJCAI, BigData, etc.). His research is generously supported by Virginia Tech, NSF, DARPA, DHS, Commonwealth Cyber Initiative, 4VA, Deloitte, Amazon, and Cisco. His work has been recognized by the 24th CNSF Capitol Hill Science Exhibition, Cisco Faculty Research Award (2023), AAAI New Faculty Highlights roster (2024), Amazon-Initiative Research Award (2024), and NSF Career Award (2024).

Abstract: A common and fundamental property of real-world data is the long-tailed distribution, where the majority of examples come from a few head categories, while the rest of the examples belong to a massive number of tail categories. This data characteristic finds broad applicability across various domains, including financial fraud detection, e-commerce recommendation, scientific discovery, and rare disease diagnosis. Despite the tremendous progress that has been made, the vast majority of existing long-tailed learning work is essentially conducted in a closed-world environment with predefined domains, data distributions, and downstream tasks. A natural and fundamental research question largely remains nascent: How can we enable open-world long-tailed learning (OpenLT) in which data is collected from heterogeneous sources with varying data distribution and the patterns of interest are evolving and open-ended? In this talk, I will discuss our group’s recent work on 1) OpenLT Theory – characterizing the task complexity and generalization performance of long-tailed learning, 2) OpenLT Algorithm – developing a generic computational framework for long-tailed learning with label scarcity and highly-skewed data distribution, and 3) OpenLT Application - hinging on key application domains to discuss our proposed techniques and theoretical results for open-world long-tailed learning. Finally, I will conclude this talk and share thoughts about my future research.

Towards Data-Centric Time Series Analysis

Presenter: Dongjin Song

Bio: Dongjin Song has been an Assistant Professor in the School of Computing at the University of Connecticut since Fall 2020. Previously, he was a Research Staff Member at NEC Labs America in Princeton, NJ. He earned his Ph.D. in Electrical and Computer Engineering (ECE) from the University of California, San Diego (UCSD) in 2016. His research interests include machine learning, data science, and their applications in time series data analysis and graph representation learning. His work has been published in top-tier data science and artificial intelligence venues, including NeurIPS, ICML, ICLR, KDD, ICDM, SDM, AAAI, IJCAI, CVPR, and ICCV. Three of his papers have been recognized as the most influential papers by paperdigest.org. He serves as an Associate Editor for Pattern Recognition and Neurocomputing, and has contributed as an Area Chair or Senior Program Committee Member for conferences such as AAAI, IJCAI, ICDM, and CIKM. He has also co-organized the AI for Time Series (AI4TS) Workshop at IJCAI, AAAI, ICDM, and SDM, as well as the MiLeTS workshops at KDD. He won with the prestigious NSF CAREER Award and the Frontiers of Science Award (FSA) in 2024.

Abstract: The increasing ubiquity of time series data across various domains, from healthcare to finance, calls for innovative approaches to analyze and interpret complex temporal dynamics. This talk explores the emerging field of data-centric methodologies in time series analysis, drawing insights from two recent advances: Rank Supervised Contrastive Learning (RankSCL) and Semantic Space Informed Prompt Learning with Large Language Models (S2IP-LLM). RankSCL enhances time series classification by integrating rank-aware contrastive learning, assigning differential importance to positive pairs and employing targeted augmentation to enrich class-specific boundaries. S2IP-LLM, on the other hand, introduces a novel alignment of pre-trained semantic spaces with time series embeddings, leveraging large language models to achieve superior forecasting performance by augmenting time series embeddings based on pre-trained word token embeddings. Together, these approaches highlight the potential of integrating advanced representation learning and data augmentation techniques to push the boundaries of predictive performance and interpretability in time series tasks. The talk will showcase empirical results on benchmark datasets, emphasizing how data-centric innovations enable robust and generalizable solutions to real-world time series challenges.

Scientific Equation Discovery via Programming with Large Language Models

Presenter: Chandan Reddy

Bio: Chandan Reddy is a Professor in the Department of Computer Science at Virginia Tech. He received his Ph.D. from Cornell University and his M.S. from Michigan State University. His primary research interests include Machine Learning and Natural Language Processing, with applications in Healthcare, Software, E-commerce, and Human Resource Management. Dr. Reddy's research has been funded by organizations such as the NSF, NIH, DOE, DOT, and various industries. He has authored over 190 peer-reviewed articles in leading conferences and journals. He received several awards for his research work including the Best Application Paper Award at the ACM SIGKDD conference in 2010, the Best Poster Award at the IEEE VAST conference in 2014, and the Best Student Paper Award at the IEEE ICDM conference in 2016. He was also a finalist in the INFORMS Franz Edelman Award Competition in 2011. Dr. Reddy serves (or has served) on the editorial boards of journals such as ACM TKDD, ACM TIST, NPJ AI, and IEEE Big Data. He is a Senior Member of the IEEE and a Distinguished Member of the ACM. More information about his work is given at https://creddy.net.

Abstract: Mathematical equation discovery is a crucial aspect of computational scientific discovery, traditionally approached through symbolic regression (SR) methods that focus mainly on data-driven equation search. Current approaches often struggle to fully leverage the rich domain-specific knowledge that scientists typically rely on. We present LLM-SR, an iterative approach that combines the power of large language models (LLMs) with evolutionary program search and data-driven optimization to discover scientific equations more effectively and efficiently while incorporating scientific prior knowledge. LLM-SR integrates several key aspects of the discovery process, namely, scientific knowledge representation and reasoning (via LLMs’ prompting and prior knowledge), hypothesis generation (equation skeleton proposals produced by LLMs), data-driven evaluation and optimization, and evolutionary search for iterative refinement. Through this integration, our approach discovers interpretable and physically meaningful equations while ensuring efficient exploration of the equation search space and generalization to out-of-domain data. We will demonstrate LLM-SR’s effectiveness across various scientific domains - nonlinear oscillators, bacterial growth, and material stress behavior. This work not only improves the accuracy and interpretability of discovered equations but also enhances the efficiency of the search process by leveraging scientific prior knowledge.