Neuro-symbolic integration for graph data

Manfred Jaeger

Associate Professor at Aalborg University

Learning and reasoning with graph and network data has developed as an area of increasing importance over recent years. Social networks, knowledge graphs, sensor and traffic networks are only some of the examples where graph-structured data arises in important applications. Much of the attention currently focuses on graph neural networks (GNNs) as the technology for solving the challenges posed by this kind of data. While often very powerful in terms of scalability and predictive performance, graph neural networks suffer from the same drawbacks as other deep learning methods: lack of interpretability, limited support for the integration of prior domain knowledge, lack of robustness, and the inability to support more flexible reasoning than performing a specific task of prediction or synthetic graph generation. The field of statistical relational learning (SRL) has been concerned with learning and reasoning with graph and network data for over 20 years. Here the use of logic-based, symbolic representations and inference techniques, probabilistic graphical models, and relational database technology supports the construction of interpretable models via a combination of expert knowledge and machine learning, as well as a wide range of inference tasks, such as prediction and (most probable) explanations for varying and incomplete amounts of input data. On the other hand, SRL techniques lag behind GNNs in terms of scalability and predictive power in scenarios where the availability of extensive training data enables the training of the highly parameterized GNN models. The objective of this research visit is to develop combinations of symbolic SRL representations with graph neural architectures in order to harness the respective strengths of GNN and SRL techniques in a coherent manner.

Keywords: Machine Learning, Graph data, Neuro-symbolic integration

Scientific area: Artificial Intelligence

Bio: Manfred Jaeger is an associated professor in computer science at Aalborg University. His research is centered on the combination of probabilistic and logic-symbolic approaches to knowledge representation, reasoning, and learning in artificial intelligence and machine learning. His work covers the development of conceptual and theoretical foundations, as well as their implementation and evaluation in software prototypes. Manfred obtained his PhD in 1995 from the University of the Saarland with work on probabilistic logics for reasoning with statistical and subjective probabilities. He subsequently focused on the integration of probabilistic graphical models with logical and relational modeling approaches. Manfred has published over 75 research papers in AI, machine learning, statistics, and theoretical computer science. He has served as associate editor for the Journal of Artificial Intelligence Research, AI Journal, and the Machine Learning Journal.

Visiting period: 14/03/2022-15/05/2022 and 08/04/2024 – 19/04/2024 at DISI-Dipartimento di Ingegneria e Scienza dell’Informazione, Trento University