The 10th Mining and Learning from Time Series Workshop: From Classical Methods to LLMs
(KDD MILETS Workshop 2024)
Held in conjunction with KDD'24
Aug 26th, 2024
Introduction
Time series data are ubiquitous. In domains as diverse as finance, retail, entertainment, transportation and health care, we observe a fundamental shift away from parsimonious, infrequent measurement to nearly continuous monitoring and recording. Recent
advances in diverse sensing technologies, ranging from remote sensors to wearables and social sensing, are generating a rapid growth in the size and complexity of time series archives. Thus, although time series analysis has been studied extensively, its importance only continues to grow. What is more, modern time series data pose significant challenges to existing techniques (e.g., irregular sampling in hospital records and spatiotemporal structure in
climate data). Finally, time series mining research is challenging and rewarding because it bridges a variety of disciplines and demands interdisciplinary solutions. Now is the time to discuss the next generation of temporal mining
algorithms.
The focus of MiLeTS workshop is to synergize the research in this area and discuss both new and open problems in time series analysis and mining. The solutions to these problems may be algorithmic, theoretical, statistical,
or systems-based in nature. Further, MiLeTS emphasizes applications to high impact or relatively new domains, including but not limited to biology, health and medicine, climate and weather, road traffic, astronomy, and energy.
The MiLeTS workshop will discuss a broad variety of topics related to time series, including:
- Time series forecasting and prediction using classical approaches
- Time series forecasting and prediction using LLMs
- Time series pattern mining and detection, representation, searching and indexing, classification, clustering, prediction, forecasting, and rule mining.
- Time series with special structure: spatiotemporal (e.g., traffic speeds at different locations), relational (e.g., patients with similar diseases), hierarchical, etc.
- Time series with sparse or irregular sampling, missing values at and not at random, and special types of measurement noise or bias.
- Time series that are multivariate, high-dimensional, heterogeneous, etc., or that possess other atypical properties.
- Time series analysis using less traditional approaches, such as deep learning and subspace clustering.
- Privacy preserving time series mining and learning.
- Online, high-speed learning and mining from streaming time series.
- Uncertain time series mining.
- Applications to high impact or relatively new time series domains, such as health and medicine, road traffic, and air quality.
- New, open, or unsolved problems in time series analysis and mining.
Schedule
09:00-09:10 Opening remarks
09:10-10:00 Keynote Talk
- New Methods for Robust Time-Series Analysis , Prof. B. Aditya Prakash
10:00-10:20 Contributed talks
- Multi-output Ensembles for Multi-step Ahead Forecasting, Vitor Cerqueira, Carlos Soares and Luis Torgo.
- SCoherency Loss for Hierarchical Time Series Forecasting, Rares Cristian, Michael Hensgen, Pavithra Harsha, Georgia Perakis and Brian Quanz.
10:20-11:05 Poster Session + Coffee Break
11:05-11:15 pm Contributed Talk
- FrENTs: Probabilistic Forecasting by Frequency Enhanced Neural Processes, Sayan Sinha, Harshavardhan Kamarthi, Alexander Rodríguez, Joy Arulraj and B. Aditya Prakash
11:15-12:05 Keynote Talk
- Modern Aspects of Time Series Forecasting , Dr. Bernie Wang
12:05-12:55 pm Keynote Talk
- Towards A Universal Model for Urban Spatio-Temporal Prediction: A Knowledge-informed Generative Learning Perspective., Dr. Jingtao Ding
12:55-1:00 pm Concluding Remarks
Speakers
Modern Aspects of Time Series Forecasting
Time series forecasting plays a pivotal role in automating and optimizing business processes. In recent years, we have witnessed a significant paradigm shift from traditional model- and assumption-based methods to data-driven, fully automated approaches. This transformation is largely driven by the increasing availability of large, diverse, and complex time series data sources, which in turn present a new set of challenges. In this keynote, we will embark on a comprehensive exploration of probabilistic time series forecasting, beginning with deep models that seamlessly blend the expressive power of neural networks with the data efficiency of classical time series methods. We will then delve into AutoML for time series and discuss the emerging trend of foundation models in this field. Finally, we will explore the practical aspects of building forecasting systems using the AWS Forecasting stack, including tools such as AutoGluon, GluonTS, SageMaker (Canvas), and Amazon Forecast.
Biography
Bernie Wang is a Principal Machine Learning Scientist at AWS AI Labs, where he leads a research group specializing in AutoML and large-scale time series forecasting, with applications spanning AIOps, supply chain optimization, workforce planning, and more. He holds a PhD in Computer Science from Tufts University, USA, and an MS in Computer Science from Tsinghua University, China. Bernie’s research interests encompass statistical machine learning, numerical linear algebra, and random matrix theory. His work in time series forecasting covers a broad spectrum, from practical applications to the development of theoretical foundations.
New Methods for Robust Time-series Analysis
Sequence modeling at scale is an important problem for many domains like traffic prediction, behavior modeling, network analytics, observability and disease forecasting. Improvements in ML models for sequence prediction have led to improvements in tasks across multiple disciplines. However, most time series prediction models only focus on producing accurate so-called ‘point’ predictions and do not prioritize handling uncertainty and calibrating the (complex, possibly multimodal) predictive distribution itself. This is problematic as providing a trustworthy and reliable estimate is important for real world applications and more generally dealing with uncertain cases during inference. This problem is exacerbated in spatio-temporal forecasting. In this talk, we discuss some of our recent work in advancing the field of multi-variate forecasting with uncertainty and calibration by bridging deep sequential models with Gaussian processes and also using neural models to handle data revisions in real-time. We will also talk about our recent work on enhancing generalizability for time-series analysis by using LLMs and developing principled multi-domain pre-trained models and adapting to industrial use cases.
Biography:
B. Aditya Prakash is an Associate Professor in the College of Computing at the Georgia Institute of Technology (“Georgia Tech”). He received a Ph.D. from the Computer Science Department at Carnegie Mellon University, and a B.Tech (in CS) from the Indian Institute of Technology (IIT) – Bombay. He has published one book, more than 100 papers in major venues, holds two U.S. patents and has given several tutorials at leading conferences. His work has also received multiple best-paper/best-of-conference selections and travel awards. His research interests include Data Science, Machine Learning and AI, with emphasis on big-data problems in large real-world networks and time-series, with applications to health, urban computing, security and the Web. Tools developed by his group have been in use in many places including ORNL, the CDC, Dow Inc and Walmart. He has received several awards such as multiple Facebook Faculty Awards, the NSF CAREER award, and was named as one of ‘AI Ten to Watch’ by the IEEE. His work has also won awards in multiple data science challenges (e.g the Catalyst COVID19 Symptom Challenge) and been highlighted by several media outlets/popular press like FiveThirtyEight.com. He is also a core-faculty at the Center for Machine Learning (ML@GT) and the Institute for Data Engineering and Science (IDEaS) at Georgia Tech.
Jingtao Ding
Postdoctoral Research Fellow
CDepartment of Electronic Engineering,
Tsinghua University
Towards A Universal Model for Urban Spatio-Temporal Prediction: A Knowledge- informed Generative Learning Perspective.
Urban spatio-temporal prediction is essential for effective decision-making in areas such as transportation management and resource optimization. While cutting-edge prediction models have advanced, they often address specific problems without generalizing across different domains. Generative models, particularly Generative Pre-trained Transformers (GPT) and Diffusion Probabilistic Models, have demonstrated significant capabilities in text and video applications. However, their potential for urban spatio-temporal prediction remains largely unexplored. This talk will introduce a versatile generative learning framework that leverages urban spatio-temporal knowledge for universal modeling. The presentation will cover 1) the construction of an urban knowledge graph to capture semantic relationships among spatial regions, temporal context, and urban flows; 2) the design of physics-informed diffusion models to characterize intrinsic mechanisms behind spatio-temporal data; and 3) the development of knowledge-guided prompt learning to create a one-for-all solution applicable across diverse scenarios.
Biography:
Jingtao Ding
Dr. Ding (https://fi.ee.tsinghua.edu.cn/~dingjingtao) is currently a Postdoctoral
Research Fellow in the Department of Electronic Engineering at Tsinghua University. He obtained his Ph.D. Degree and Bachelor’s Degree from the same department in 2020 and 2015, respectively. During the years 2020 and 2022, he worked as an applied researcher at Wechat, the largest SNS company in China. His current research focuses primarily on generative modeling of complex human behavior and dynamics in urban system. He has over 40 publications in top-tier venues on data science and machine learning, such as KDD, NeurIPS, WWW, ICLR, IJCAI, IEEE TKDE, ACM TOIS, etc. He serves as the academic editor for PLOS Complex Systems and PC member for several top-tier conferences. His work on understanding user behaviors in the complex network of recommender system won the Best Poster Paper Award in WWW 2018.
Accepted Papers
Understanding Fear and Beyond in Neuronal Networks with Tensor and Graph Methods: An Interdisciplinary End-to-End Data Science Approach . Jordan Steinhauser, Edward Korzus and Evangelos Papalexakis.
FrENTs: Probabilistic Forecasting by Frequency Enhanced Neural Processes. Sayan Sinha, Harshavardhan Kamarthi, Alexander Rodríguez, Joy Arulraj and B. Aditya Prakash.
Fine-grained Attention in Hierarchical Transformers for Tabular Time-series. Raphael Azorin, Zied Ben Houidi, Massimo Gallo, Alessandro Finamore and Pietro Michiardi.
Time series forecasting with high stakes: A field study of the air cargo industry. Naman Shukla and Abhinav Garg.
SCoherency Loss for Hierarchical Time Series Forecasting. Rares Cristian, Michael Hensgen, Pavithra Harsha, Georgia Perakis and Brian Quanz.
Towards Foundation Auto-Encoders for Time-Series Anomaly Detection. Gastón García González, Pedro Casas, Emilio Martínez and Alicia Fernández.
Multi-output Ensembles for Multi-step Ahead Forecasting. Vitor Cerqueira, Carlos Soares and Luis Torgo.
Attention-guided Neural Differential Equations Framework for Missingness in Time series Classification. Yongkyung Oh, Dongyoung Lim, Sungil Kim and Alex Bui.
Call for Papers
Submissions should follow the SIGKDD formatting requirements (unless otherwise stated) and will be evaluated using the SIGKDD Research Track evaluation criteria. Preference will be given to papers that are reproducible, and authors are encouraged to share their data and code publicly whenever possible. Submissions are limited to be no more than 9 pages (suggested 4-8 pages), including references (all in a single pdf). All submissions must be in pdf format using the KDD main conference paper template (see: https://kdd2024.kdd.org/research-track-call-for-papers/ ). Submissions will be managed via the MiLeTS 2024 EasyChair website: https://easychair.org/conferences/?conf=milets2024.
Note on open problem submissions: In order to promote new and innovative research on time series, we plan to accept a small number of high quality manuscripts describing open problems in time series analysis and mining. Such papers should provide a clear, detailed description and analysis of a new or open problem that poses a significant challenge to existing techniques, as well as a thorough empirical investigation demonstrating that current methods are insufficient.
The review process is single-round and double-blind (submission files have to be anonymized). Concurrent submissions to other journals and conferences are acceptable. Accepted papers will be presented as posters during the workshop and list on the website (non-archival/without proceedings). Besides, a small number of accepted papers will be selected to be presented as contributed talks.
Any questions may be directed to the workshop e-mail address: kdd.milets@gmail.com.
Download Call for Papers here.
Key Dates
Paper Submission Deadline: May 28th June 21st, 2024, 11:59PM Alofi Time
Author Notification: June 28th July 21st, 2024, 11:59PM Alofi Time
Camera Ready Version: August 2nd, 2024
Workshop: August 26th, 2024
