Tutorials

Time series data is widely applied in various fields such as transportation, healthcare, and energy. This report focuses on the time series foundation models for forecasting, anomaly detection, and classification tasks, systematically introducing several key elements of their construction, such as the selection and construction of pre-training data, the design of generalizable model architectures and training strategies, and presents the current mainstream methods and models from these aspects. Through this report, the audience will gain a comprehensive understanding of the technical system and development trends of time series foundation models, providing references for subsequent research and practical applications. As a result, research on time series methods has become crucial. To advance progress in this area, we propose OpenTS, an automated benchmarking framework for time series forecasting and anomaly detection methods. OpenTS addresses current research challenges, including insufficient coverage of data domains, bias toward traditional methods, and inconsistent and inflexible processes. We improve data domain coverage by incorporating datasets from 10 different fields and provide time series characterization to ensure the comprehensiveness of these datasets. Additionally, OpenTS supports the integration of various methods, including statistical learning, machine learning, and deep learning approaches, and offers multiple evaluation strategies and metrics to ensure comprehensive assessment of different methods. Moreover, with the recent rising of time series foundation models, OpenTS also includes the recently proposed methods in different forecasting strategies, including zero-shot, few-shot, and full-shot, thereby facilitating more thorough evaluations.

Environmental, Social, and Governance (ESG) data has become a critical foundation for sustainable finance, corporate accountability, and responsible AI, yet remains fragmented, unstructured, and difficult to manage. This tutorial presents a unified perspective on AI-enhanced ESG data management, introducing a lifecycle framework that integrates large language models (LLMs) and modern data systems to support transparency, automation, and trust in ESG analytics. Building on recent advances in responsible data management and human-centric AI, we demonstrate how LLMs can extract, validate, and standardize ESG information across diverse sources and global regulatory frameworks (e.g., GRI, SASB, CSRD, ISSB), enabling efficient compliance checking, continuous monitoring, and real-time stakeholder engagement. Through theoretical foundations, system architecture, and industry case studies, the tutorial bridges AI, data management, and sustainability governance, offering participants practical insights into designing next-generation ESG intelligence platforms for trustworthy and socially beneficial AI applications.

As the volume and ubiquity of graphs increase, a compact graph representation becomes essential for enabling efficient storage, transfer, and processing of graphs. Given a graph, the graph summarization problem seeks a compact representation that comprises a summary graph and corrections, allowing for the exact recreation of the original graph from the representation. Studies in this field aim to explore general, queryable compressed storage structures for graph data, which are considered a potential solution for graph processing tasks in memory-constrained scenarios. In this tutorial, we first highlight the importance of graph summarization in a variety of applications and the unique challenges that need to be addressed. Subsequently, we provide an overview of the existing methods for the graph summarization problem and the research on its various variants. Finally, we discuss the future research directions in this important and growing research area.

Longitudinal data are generated by many sensors that are being massively deployed across wide range of systems. This yields very large amounts of multi-dimensional and possibly heterogeneous time series data that must be processed to identify changes in the system behavior, to detect events, and to predict the future values in order to control the system states. The temporal dimension of longitudinal data prohibits processing all samples at once. It necessitates using statistical filtering techniques, especially when the processing needs to be performed in real-time, and when the outputs values need to be generated at the same rate as the input samples are arriving. The filtering must capture the important temporal patterns across many time-scales. Another challenge is how to effectively process multiple time-series to capture their spatial dependencies without excessively increasing the computational complexity. The traditional model-based approaches often rely on Gaussianity of samples, and their well-defined correlation patterns. However, in many scenarios, the samples are non-Gaussian, which requires adopting non-linear modeling techniques. When even non-linear models fail to achieve an acceptable accuracy, the sample-based universal model-free methods can be considered. These methods usually combine various signal decompositions with deep learning filtering. They are effective for whole classes of input signals rather than being tailored to one specific type of the signal unlike statistical filtering methods. The drawback of the model-free methods is that they often require time consuming and expensive training. They are also numerically much more complex, which can be prohibitive in some practical implementations. The objectives of this tutorial are as follows. (1) Describe the goals, challenges and opportunities in processing the real-world longitudinal data; (2) Explain the pros and cons of using traditional adaptive filtering techniques vs. using modern deep learning architectures; (3) Explore implementation constraints in developing practical applications assuming the speech applications as an example; and, (4) Summarize the recent literature, and outline the open research problems.