About the project
The overarching goal of this PhD project is to develop a general interpretable framework for forecasting on dynamic graph-structured data. More specifically, we will extract a temporal logical program from a trained Neural Network (NN) model that takes as input Dynamic Knowledge Graphs. The envisioned approach falls under the category of neuro-symbolic Artificial Intelligence (AI), combining Deep Learning with traditional logic-based AI, and builds upon recent connections between Graph Neural Networks (GNNs) and logic.
Machine Learning (ML), in particular, NNs, is at the centre of the dramatic success of AI in recent years. However, while NNs are an incredible tool and have led to great results in many domains, they have limitations. Most notably, they are often seen as black boxes, and making NNs inherently interpretable or providing explanations for their behaviour are active fields of research. More traditional AI approaches, the so-called symbolic approaches, have complementary benefits to NNs. This led to the emergence of hybrid neuro-symbolic approaches which combine the two paradigms, mitigating their downsides and capitalising on their benefits.
There are many ways of integrating symbolic AI methods and NNs, but the combination I am interested in is extracting a logical program from a trained GNN model. The idea is to train a GNN to make predictions, extract a logical program that can make the same or approximately the same predictions, and then use the program on the inference flow. In this way I obtain a solution that is data-driven and with a good prediction accuracy (NN benefits), but that is human-interpretable and can be easily integrated with prior logical knowledge and constraints (symbolic benefits). Moreover, there exists early work extracting logical programs from trained models of a specific class of GNNs applied to static data. I plan to expand this to more complex NNs that work on temporal data.
To realise this vision, we have three objectives for this project:
- Develop a simple yet effective GNN-based architecture for forecasting on temporal graph-structured data.
- Develop a translation algorithm for extracting a logical program (in a temporal logic language) from the trained model of the network developed in Objective 1.
- Construct a hybrid framework that encompasses the training of the GNN, the translation into a logical program and the application of the program for generating predictions on new data.
Project team
- PhD Fellow Roxana Pop, University of Oslo, Department of Informatics
- Associate Professor Egor V. Kostylev, University of Oslo, Department of Informatics
- Senior Researcher Przemysław A. Wałęga, University of Oxford