Keynote Talk

Verification of Graph Learning Models and Analysis of Query Languages

Graph neural networks (GNNs) are the predominant architectures for a variety of learning tasks on graphs. Like other modern machine learning models, their success is accompanied by many fundamental concerns. Can we understand what exactly a trained model does? Can we have confidence that a trained model will not do anything bad? In this talk we will give one of the first approaches to verifying GNNs. We will work using the static analysis of techniques of database theory. Many of our results work by converting the GNN into a logic, and then showing that we can analyze the logic. This translation is useful not only for verification, but for explanation of the behavior of GNNs.

The work reported here is joint with Chia-Hsuan Lu, Boris Motik, and Tony Tan (Decidability of Graph Neural Networks via Logical Characterizations), it is closely related to work of Barcelo, Kostylev, Monet, Reutter, and Silva from ICLR 2020 (The Logical Expressiveness of Graph Neural Networks), and also to prior work on analysis of logics on graphs with arithmetic (On two-variable guarded fragment logic with expressive local Presburger constraints; Two Variable Logic with Ultimately Periodic Counting). The talk will be in the same spirit as the AMW summer school course on analysis of GNNs via query languages: but there will be no dependency between the two, and no overlap in the results.

Bio

Michael Benedikt is a professor at Oxford University’s computer science department, and a fellow of University College Oxford. He came to Oxford after a decade in US industrial research laboratories, including a position as Distinguished Member of Technical Staff at Bell Laboratories. He has worked extensively in computational logic, finite model theory, verification, and data management, and specializes in the interaction between these topics. He has been a keynote in past meetings on mathematical logic, computational logic, description logics, and databases. He has co-authored papers receiving best paper awards and test-of-time awards in major conferences within databases and theoretical computer science, and he has served as the program chair of both the ACM Principles of Database Systems conference (2012) and the International Conference on Database Theory (2017). He currently holds an Established Career Fellowship from the UK’s Engineering and Physical Science’s Research Council, and serves on the steering committees for the Association for Symbolic Logic, the European Association for Theoretical Computer Science, and the International Conference on Database Theory.

Keynote Talk

Integrating Semantics with Symbolic AI: The Path to Interpretable Hybrid AI Systems

Bio

Prof. Dr. Maria-Esther Vidal is the head of the Scientific Data Management Research Group at TIB and a member of the L3S Research Centre at the University of Hannover; she is also a full professor (on-leave) at Universidad Simón Bolívar (USB) Venezuela. Her interests include Big data and knowledge management, knowledge representation, and semantic web.  She has published more than 170 peer-reviewed papers in Semantic Web, Databases, Bioinformatics, and Artificial Intelligence. She has co-authored one monograph, and co-edited books and journal special issues. She is part of various editorial boards (e.g., JWS, JDIQ), and has been the general chair, co-chair, senior member, and reviewer of several scientific events and journals (e.g., ESWC, AAAI, AMW, WWW, KDE). She is leading data management tasks in the EU H2020 projects iASiS, BigMedylitics, and QualiChain, and has participated in BigDataEurope, BigDataOcean; she is a supervisor of MSCA-ETN projects WDAqua and NoBIAS. She has been a visiting professor in different universities (e.g., Uni Maryland, UPM Madrid, UPC, KIT Karlsruhe, Uni Nantes). In the past, she has participated in international projects (e.g., FP7, NSF, AECI), and led industrial data integration projects for more than 10 years (e.g., Bell South, Telefonica).

Keynote Talk

Learning in Query Optimization over Knowledge Graphs: From Adaptive Techniques to Neuro-Symbolic Optimizers … And Back?

Query optimization has traditionally relied on the optimize-then-execute paradigm, which, while effective in static settings, faces significant limitations in dynamic and complex environments such as knowledge graphs on the web. In this keynote, I will explore how early database adaptive techniques provided the first steps toward online learning query optimization over knowledge graphs, allowing systems to adjust during execution. I will present results on when adaptivity enhances performance in knowledge graphs and when it falls short.

Following this, I will discuss the rise of neuro-symbolic optimizers—recent innovations that combine machine learning with symbolic processing (e.g., rules, statistics, data summaries, etc.). Therefore, neuro-symbolic optimizers promise to deliver more accurate results than traditional optimizers, especially in the presence of increasingly complex workloads and scenarios. This raises a fundamental question: Can neuro-symbolic optimizers eliminate the need for adaptive techniques, or will systems still require the flexibility to adapt during execution? I will conclude with an exploration of this open question, the challenges of embedding machine learning into query optimization, and the interplay of machine learning and adaptivity, with a special focus on knowledge graphs and their unique requirements.

Bio

Maribel Acosta is the Professor of Data Engineering at the TUM Campus Heilbronn since August 2023. Maribel Acosta studied Computer Science at Universidad Simon Bolivar, Venezuela. From 2012 to 2017, she was a research assistant at the Karlsruhe Institute of Technology (KIT), where she received her doctorate. She then worked as a postdoc and deputy professor at KIT until 2020. Afterward, she was appointed as the professor for Databases and Information Systems at the Ruhr-University Bochum until July 2023. She is actively involved in the scientific communities on Data Management and Artificial Intelligence. Her work has received several “Best Paper Awards” and she serves as chair and reviewer for renowned conferences. Besides research, Maribel Acosta has many years of teaching experience in Databases, Big Data, and Knowledge Graphs and has received two “Best Teaching Awards”. Maribel Acosta investigates techniques for managing knowledge graphs. Her contributions include efficient solutions for querying knowledge graphs while providing high-quality answers.

Keynote Talk

Data/AI Assemblages of Inequality: Feminist and Decolonial Perspectives from/for Latin America

Data and AI are at the center of global attention, with widespread promises and narratives about fostering economic growth and increasing productivity across all spheres of social life. These narratives are rapidly adopted by governments, organizations, and companies in our region and across the majority world. However, it is crucial to critically examine what this means for us, academics and practitioners, particularly as rising inequalities are being reproduced and amplified by data/AI assemblages globally. By data/AI assemblages, we refer to the complex and dynamic configurations of power relations, governance structures, material infrastructures, and cultural practices that shape the current data and AI ecosystem. These assemblages are never neutral; they reflect and reinforce existing and historical power imbalances—at the level of global governance, within data/AI ecosystems, and through the AI systems that increasingly shape our social and natural worlds.

Bio

Paola Ricaurte is an associate professor at the Department of Media and Digital Culture at Tecnológico de Monterrey and a faculty associate at the Berkman Klein Center for Internet & Society at Harvard University. Together with Nick Couldry and Ulises Mejías, she co-founded Tierra Común, a network of academics, practitioners and activists interested in decoloniality and data. She participates in several expert committees, such as the Global Partnership for Artificial Intelligence (GPAI), the Global Index on Responsible AI and the Expert Group for the implementation of the UNESCO Recommendation on the Ethics of AI. She is a member of the A Plus Alliance for Inclusive Algorithms and coordinates the Latin American and Caribbean hub of the Feminist AI Research Network, from where she promotes the development of Feminist AI. In addition to her academic work, she participates in civil society initiatives to promote the development of public interest technologies.

Keynote Talk

From Strings to Things and Back: Putting Logic-based Reasoning into Action in a Central Bank

The talk will present a story of the development of new methodologies, tools, and systems in the Datalog+/- realm, conducted through multiple projects with the Joint Knowledge Graph Lab, which includes the Applied Research Team of the Bank of Italy and the KG Lab of TU Wien. Specifically, a concise overview of recent advancements in implementing scalable Datalog+/- fragments in the Vadalog System will be interleaved with real-world applications in the central banking domain. These applications cover areas such as banking supervision, economic research, data privacy, and large-scale reasoning in complex financial Knowledge Graphs.

Given the recent emergence of Large Language Models, the talk will be an opportunity to discuss the evolving role of logic-based reasoning systems based on Knowledge Graphs (“things, not strings”, as Google stated in 2012). These systems, such as Vadalog, have a unique potential to balance the data-driven, deterministic, and explainable nature of deductive reasoning with the flexibility of modern natural language processing (“strings, not things,” as we might say), which often fails, especially when addressing enterprise questions requiring data and knowledge not available in established public databases. Finally, the talk will offer a chance to comment on the challenges an R&D team faces when spinning off production-ready projects from core database research, which, despite the difficulties, often results in high user satisfaction due to the unique qualities of logic-based approaches, such as explainability, scalability, and the creation of valuable enterprise data assets.

Bio

Luigi Bellomarini is a manager, researcher, and software engineer, currently serving as the Head of the Applied Research Team (the R&D Unit, in the IT Directorate General) of the Central Bank of Italy. He is also a Contract Professor of Database Systems at Roma Tre University and a Guest Lecturer on Knowledge Graphs at the University of Oxford. Additionally, he is affiliated with the Center for AI and Machine Learning (CAIML) at TU Wien, an inter-faculty center that brings together researchers in AI and ML, where he leads the Industrial Advisory Board of the SIG Knowledge Graphs group. He holds a Ph.D. in Computer Science and Engineering from the University of Roma Tre and has been a Visiting Scientific Collaborator at the Department of Computer Science at the University of Oxford.

He has a track record of leading R&D projects in both academic and industrial contexts, with a focus on translating theory into practice, particularly in economic, financial, and statistical applications. His research interests encompass a wide range of topics in databases and artificial intelligence, including knowledge-based and logic-based reasoning, neuro-symbolic methods, knowledge graphs, database theory, big data, data integration, and data exchange. He has published his research in leading venues in his fields, including VLDB, IJCAI, ICDE, EDBT, the Journal of Data Semantics, and the Journal of Information Systems.