Project Summary:

Computer and sensor networks, link structure of websites, social networks, data organization and the flow of computation in general are a few examples of relational systems that can be modelled as graph representations in the context of computer science. In particular, graphs and graph theory has been proved as a powerful technique to detect cyber and physical attacks, abnormal user behaviour or web spam.

The success of graph theory in many of this application in comparison with other machine learning technologies, where only the raw data is used, is justified due to the reduced limited of examples for supervising the learning process. Thus, the exploitation of the underlying structure of the data helps mitigate the lack of training data. Attempting to better use this additional source of information, initial machine learning approaches to cope with graph structured data made use of a preprocessing stage which maps the graph to a simpler representation, such as a matrix representation or a vector of reals. However, this process implies a compression or ‘squash’  of the graph structure, where information may be lost or not fully extracted.

Recently techniques combining graph representation with neural networks have been proposed, such as Graph Neural networks GNN, Relational Neural Networks (RelNNs) and Probability Mapping Graph Self Organizing Maps (PM-GraphSOMs).  These techniques have shown their potential to represent any type of general graph (directional, bidirectional, cyclic, etc..) and their attributes, and deal directly with graph structured information. However, these techniques do not fully exploit recent advances in deep learning in their architectures. They are also limited to static domains, where the input graph cannot change over time. In this project we will investigate techniques for addressing these current limitations.

Contact Details:

Jesus Martinez del Rincon


Telephone:+44 (0)28 9097 5626