Temporal Graph Learning Workshop
@ NeurIPS 2023

Speaker Bio: Daniele Zambon is a postdoctoral researcher at the Swiss AI Lab IDSIA at the Università della Svizzera italiana (Switzerland). He received his Ph.D. degree from the Università della Svizzera italiana with a thesis on anomaly and change detection in sequences of graphs. He holds Master's and Bachelor's degrees in mathematics from the Università degli Studi di Milano (Italy). Daniele has been a visiting researcher/intern at the University of Florida (US), the University of Exeter (UK), and STMicroelectronics (Italy). His main research interests encompass graph representation learning and learning in non-stationary environments. He is a member of the IEEE CIS Task Force on Learning for Graphs and has co-organized special sessions and tutorials on deep learning and graph data. He regularly publishes and reviews for top-tier journals and conferences in the field, including IEEE TNNLS, IEEE TSP, IEEE TPAMI, JMLR, NeurIPS, ICLR, ICML, and CVPR. 

Talk Title: Graph Deep Learning for Time Series Processing

Graph-based deep learning methods have become popular tools to process collections of correlated time series.  Differently from traditional multivariate forecasting methods, neural graph-based predictors take advantage of pairwise relationships by conditioning forecasts on a (possibly dynamic) graph spanning the time series collection.  The conditioning can take the form of an architectural inductive bias on the neural forecasting architecture, resulting in a family of deep learning models called spatiotemporal graph neural networks.  In this talk I will discuss a comprehensive methodological framework for time series analysis based on graph deep learning techniques. In particular, I will provide some design principles for graph-based predictive models, methods to assess the quality of their predictions, and an overview of the associated challenges.

Speaker Bio: Kelsey Allen is currently a Senior Research Scientist at DeepMind. She received her PhD from MIT in the Computational Cognitive Science group, and her BSc from the University of British Columbia in physics. Her work has received awards including the international Glushko award for best dissertation in cognitive science, a best paper award from Robotics: Science and Systems (R:SS), and an NSERC PhD fellowship. Spanning robotics, machine learning, and cognitive science, her work aims to elucidate the mechanisms that give rise to adaptive and efficient learning in humans and machines, especially in the domain of physical problem-solving.

Talk Title: Simulation from states and sensors

“Intuitive physics”, or the ability to imagine how the future of physical systems will unfold, is a hallmark of human intelligence. This capacity supports many complex behaviors in humans including planning, problem-solving, and tool creation. In this talk, I will describe some of our work aiming to learn physical simulators from data in order to support these downstream behaviors. The first half of the talk will focus on what we can learn from state-based data. I will present a new type of graph neural network for learning realistic rigid body simulation by taking inspiration from classic approaches in graphics. I will show that these learned simulators can capture rigid body dynamics with unprecedented accuracy, including modeling real world dynamics significantly better than system identification with analytic simulators from robotics. The second half of the talk will focus on what we can learn from sensor data. I will present Visual Particle Dynamics, a method which connects neural radiance fields with graph neural networks to enable learning directly from RGB-D data. Unlike existing 2D video prediction models, we show that VPD's 3D structure enables scene editing and long-term predictions.

Speaker Bio: Ingo Scholtes is a Full Professor for Machine Learning in Complex Networks at the Center for Artificial Intelligence and Data Science of Julius-Maximilians-Universität Würzburg, Germany as well as SNSF Professor for Data Analytics at the Department of Computer Science at the University of Zürich, Switzerland. He has a background in computer science and mathematics and obtained his doctorate degree from the University of Trier, Germany. At CERN, he developed a large-scale data distribution system, which is currently used to monitor particle collision data from the ATLAS detector. After finishing his doctorate degree, he was a postdoctoral researcher at the interdisciplinary Chair of Systems Design at ETH Zürich from 2011 till 2016. In 2016 he held an interim professorship for Applied Computer Science at the Karlsruhe Institute of Technology, Germany. In 2017 he returned to ETH Zürich as a senior assistant and lecturer. In 2019 he was appointed Full Professor at the University of Wuppertal. Since 2021 he holds the Chair of Computer Science XV - Machine Learning for Complex Networks at Julius-Maximilians-Universität Würzburg, Germany.

Talk title: De Bruijn Goes Neural: Towards Causality-Aware Graph Neural Networks for Time Series Data

Graph Neural Networks (GNNs) have become a cornerstone for the application of deep learning to data on complex networks. However, we increasingly have access to time-resolved data that not only capture which nodes are connected to each other, but also when and in which temporal order those connections occur. A number of works have shown how the timing and ordering of links shapes the causal topology of networked systems, i.e. which nodes can possibly influence each other over time. Moreover, higher-order network models have been developed that allow us to model patterns in the resulting causal topology. Building on these works, we introduce De Bruijn Graph Neural Networks (DBGNNs), a novel time-aware graph neural network architecture for time-resolved data on dynamic graphs. Our approach accounts for temporal-topological patterns that unfold via causal walks, i.e. temporally ordered sequences of links by which nodes can influence each other over time. This enables us to learn patterns in the causal topology of time series data on complex networks, which facilitates to address learning tasks in temporal graphs.

Speaker Bio: I'm an assistant professor in the Department of Computer Science at Yale University. My research focus includes algorithms for graph neural networks, geometric embeddings and explainable models. I am the author of many widely used GNN algorithms such as GraphSAGE, PinSAGE and GNNExplainer. In addition, I have worked on a variety of applications of graph learning in physical simulations, social networks, knowledge graphs and biology. I developed the first billion-scale graph embedding services at Pinterest, and the graph-based anomaly detection algorithm at Amazon. 

Talk abstract: Temporal graphs are widely used to model dynamic systems with time-varying interactions. In real-world scenarios, the underlying mechanisms of generating future interactions in dynamic systems are typically governed by a set of recurring substructures within the graph, known as temporal motifs. Despite the success and prevalence of current temporal graph neural networks (TGNN), it remains unclear how to explain such temporal graph predictions. To address this challenge, we propose a novel approach to explain temporal graph predictions via temporal motifs. The method, called Temporal Motifs Explainer (TempME), uncovers the most pivotal temporal motifs guiding the prediction of TGNNs. Derived from the information bottleneck principle, TempME extracts the most interaction-related motifs while minimizing the amount of contained information to preserve the sparsity and succinctness of the explanation. Events in the explanations generated by TempME are verified to be more spatiotemporally correlated than those of existing approaches, providing more understandable insights. 

Speaker Bio: Marinka Zitnik is an Assistant Professor at Harvard University with appointments in the Department of Biomedical Informatics, Broad Institute of MIT and Harvard, and Harvard Data Science. Dr. Zitnik is a computer scientist studying applied machine learning with a focus on challenges in scientific discovery and medicine. Her algorithms and methods have had a tangible impact, which has garnered interests of government, academic, and industry researchers and has put new tools in the hands of practitioners. Some of her methods are used by major biomedical institutions, including Baylor College of Medicine, Karolinska Institute, Stanford Medical School, and Massachusetts General Hospital. Her work received best paper and research awards from International Society for Computational Biology, Bayer Early Excellence in Science Award, Amazon Faculty Research Award, Roche Alliance with Distinguished Scientists Award, Rising Star Award in Electrical Engineering and Computer Science (EECS), and Next Generation Recognition in Biomedicine. She is an ELLIS Scholar, a member of the Science Working Group at NASA Space Biology, co-founder of Therapeutics Data Commons, and faculty lead of the AI4Science initiative. She is also the recipient of the 2022 Young Mentor Award at Harvard Medical School, and was named Kavli Fellow 2023 by the National Academy of Sciences.

Talk title: Towards foundation models for time series

General-purpose foundation models have revolutionized deep learning, enabling the adaptation of a single model to a wide array of tasks with minimal additional training, eliminating the need for separate models for each task. While this paradigm has proven successful in vision and language domains, extending it to time series data poses significant challenges due to the diverse temporal dynamics, semantic variations, irregular sampling, system-related factors (e.g., different devices or subjects), and shifts in feature and label distributions inherent to time series data. These factors are not inherently compatible with the next-token prediction objective in large language models. In this talk, I describe our research efforts to realize crucial capabilities for time-series foundation models. We start with TF-C (NeurIPS 2022), a time-series pre-training strategy that leverages a self-supervised consistency objective, modeling both temporal and frequency representations within the time-frequency space. We then explore Raincoat (ICML 2023), the first approach for closed-set and universal domain adaptation that is robust to both feature and label shifts, allowing model transfer between source and unlabeled target domains, even in scenarios with no label overlap. To facilitate the analysis of time series model behavior, our TimeX approach (NeurIPS 2023) introduces an interpretable surrogate model, ensuring model behavior consistency, providing discrete attribution maps, and enhancing interpretability. Lastly, Raindrop (ICLR 2022) is an approach for learning from irregularly sampled multivariate time series, employing a graph neural network to capture time-varying dependencies among sensors and outperforming state-of-the-art methods in classification and temporal dynamics interpretation. Collectively, these approaches shed light on the evolving landscape of time series representation learning, offering a roadmap for future advancements in temporal learning.