Dynamical Network Analysis and Machine Learning for Computational Social Sciences

The drive towards digitalisation, and the increase in data access and information integration across diverse domains is reinforcing the effects of a networked society. Personal data and actions are now embedded into social networks leading to vast information flows taking place through online social interactions. These new modalities of social links have led to a huge increase in available data of diverse types and scopes, which pose challenges to traditional quantitative methods in the social sciences. Yet they also provide enormous opportunities to gain novel insights through the extraction and analysis of socio-economic data enhanced by social media. To achieve this, we need the development of mathematical and computational techniques based on dynamical and stochastic processes on networks which can capture the key characteristics of social dynamics.
This PhD project will investigate new mathematical and algorithmic tools for network analysis that address the challenges posed by big data for a networked society. Methodologically, the project lies at the intersection of network analysis, graph theory, dynamical systems, stochastic processes, and graph-based statistical learning.
Networks emerging from the analysis of social data sets have distinct properties that will conform our special focus of research. Specifically, we will focus on methods that can deal with directed weighted graphs; geometrically and geographically constrained networks; and graphs with multi-scale and multi-layer organisation that capture the richness of social interactions and information flows. As a first objective, we will consider the problem of scale selection within methods for multiscale graph partitioning based on random walks (i.e., Markov Stability analysis). This dynamical approach to graph partitioning extends naturally to directed, weighted, multiscale networks, making it highly applicable to social networks. We will then generalise this analysis to multilayer networks, which allow the integration of different kinds of interactions between nodes, and to hypergraphs, which allow for more than pairwise interactions between nodes. The theory of both multilayer networks and hypergraphs is still in its infancy and this project will contribute to a novel mathematical and algorithmic framework for their analysis exploiting the connection between dynamics and spectral properties. This mathematical work will be informed by and applied to models of consensus and opinion formation on graphs guiding the formulation of our problems.
Additionally, we will develop methods that merge networks and high-dimensional sample data by extending graph-based statistical learning using results from stochastic processes on graphs. We will also use tools from natural language processing when appropriate to include text from online media as high-dimensional vectors extracted from embeddings.
The developed methods will be applied to different domains in the computational social sciences. We will extend our analysis of UK and European human mobility patterns under COVID-19 (Facebook data) using dynamical networks, which has already uncovered insights into the socio-economic consequences of the pandemic. Furthermore, we will pursue an in-depth study of online opinion formation through our recently established collaboration with Prof. Barbara Pfetsch at the Weizenbaum Institute for the Networked Society and the Freie Universitaet in Berlin who studies polarisation of political groups online. Being aware of the ethical complexity of machine learning, we also aim to reflect on the ethics of our applications by evaluating the biases of our models in a transparent Bayesian framework.
EPSRC Research areas:Statistics and applied probability,Artificial intelligence technologies, Complexity science, Human communication in ICT, Information systems, Natural language processing
Themes:Mathematical sciences,Digital Economy,Information and communication tec