General Profile:

Hi, my name is Alex.
I started off studying astrophysics in Durham, but through my masters there, I realised my skills and passion laid more with the data I’d been using than with the theoretical pen-and-paper work of cosmology. I studied on the Data Science MSc course in 2020/21, which I found engaging, especially any work involving machine learning and artificial intelligence - which is why I applied for the CDT.
My background is quite oriented to the (mostly) well-behaved data from astronomy, and I’m still interested in this classic field of science, but I’m looking forward to working in some other areas. I don’t have any clear idea about what my PhD project will be yet, or even what field it’ll be in.
Outside of work I love being outdoors, hiking, cycling or whatever else. On rainy/lazy days I’m usually editing photos from the last time I was outdoors or watching a film.

Research Project Summary:

The summer project addressed whether recent GNN models could be used to generate synthetic social networks. The motivation for the summer project stemmed from the inherent privacy and security concerns that follow the distribution of social network data. We hypothesized that recent GNN models capable of producing graphs at scale (for example proteins) might be applied with success to social networks, and that the realism of these “fake” networks would be superior to currently employed rule-based models.

The research carried out found that the model GRAN (Gated Recurrent Attention Network) was highly capable of generating realistic social network structures. Using MMD (Maximum Mean Discrepancy) we show this empirically, though as GRAN is only capable of producing un-directed networks (mutual connections), GRAN is only tested against fairly simple network structures.

The PhD will likely explore this idea in greater detail, with several possible extensions. Initially this would focus on developing a novel GNN model capable of producing directed networks – current models are not capable of this – and subsequently on extending this model to include node attribution. Development of this directed-GNN (d-GNN for shorthand) will begin by building on existing architectures, but given the recent level of work on GNNs, which models will be built on is still speculative.

In social networks node and edge attributes are information about the people within the network and their relationships. If development of a model capable of producing attributed and directional social networks were successful, significant time would be devoted to ensuring that privacy is indeed preserved in the resulting networks, as there may be significant potential for over-fitting by the model. In this scenario, the model would reproduce structures and information very close to the real networks it has learnt from, in which case the networks would not be fit for distribution. How to ensure and measure privacy preservation in synthetic networks is itself a potential research topic.

The use of GNNs also has potential in dimensionality reduction and graph layout, which may be a focus of later research.