Pre-recorded presentations and slide decks are now available at the links below.
Presentation 1: Exploring the potential of the Nucleation phenomena in spin crossover systems to explain the social dynamics at play in neighbourhood formation
By: Cecile de Bezenac
Abstract: This analogy to biological and physical systems has led urban studies towards complexity science, drawing from the methods developed for statistical physics. Various urban ‘laws’ have been defined and the use of artificial agent-based models has become common place. In particular the Schelling model is said to illustrate the way different groups of people aggregate in a city and how segregation can appear. In this project we use the framework provided by this model to explore how the phenomenon of nucleation responsible for crystal growth in materials can be used to better understand the formation mechanisms of ‘places’, from an unstable configuration to the emergence of identifiable structured areas. Guided by a literature review covering both Nucleation studies in ferromagnetic systems and urban neighbourhood formation theories, we will conduct a simulation-based analysis in order to suggest a replicable methodology for analysing neighbourhoods.
Presentation 2: Two data-driven approaches to modelling the interactions between the UN’s Global Goals
By: Amirali Emami
Abstract: The Sustainable Development Goals were set by the UN in 2015 to be achieved by member countries by 2030. Each conceptual goal is broken down into multiple tangible targets, which are further broken down to quantitative indicators, giving a total of 230+ across the goals. Using these indicators as variables, correlations within and across countries can be explored and modelled to give a possible picture of how the goals interact.
In this talk I present two methodologies; the first developed by Ranganathan and Spaiser et al. and the second by researchers at the World Bank.
The first is a data-driven dynamical systems approach to modelling these interactions, inspired by machine learning-style model selection and applied to coupled differential equations. Dynamical systems models capture variations across time effectively, but when applied in this data-driven manner require a lot of compute and can only capture interactions between a few indicators at a time.
The second methodology is an application of the Product Space ideas from Economics to the development indicators, building a network in which each indicator is a node with the connections representing the strength of their synergies. The edges are built upon information gathered after benchmarking each country’s performance on the indicators using linear regression, and investigating which indicators appear to be co-achieved often. This method efficiently builds a network which captures the overall interactions between the indicators, but does so using only the latest available datapoint per country meaning that longitudinal information is lost.
The aim of this internship project will be to merge these to methodologies in a coherent manner, in order to scale the modelling to all indicators while also utilising the longitudinal information in the dataset.
The presentation recording will be available shortly.
Presentation 3: Understanding the issues and challenges in automated transcribing of phone calls for analysis in a Bus Transit System.
By: Brett Hull
Abstract: This project will seek to increase understanding of the way in a bus transit system which First Bus (our data partner) operates and how the complaints are recorded, processed and responded to.
As the majority of complaints received by First are received by telephone call, it will be beneficial to transcribe these calls into textual data. This textual data, along with other textual data such as complaints received by email, will be analysed using Natural Language Processing as part of the overall project.
This first stage of the project will focus on understanding the relevant business processes and any issues which may occur when working with call recordings. Methods to transcribe calls and comprehend the recordings to enable analytics will be investigated. The final step of this project will be to take a sample of recordings and attempt to accurately and automatically transcribe the call. This will be measured against a test set transcribed by a human.
Presentation 4: Predicting participation in weight management programmes based on psychometric and demographic data
By: Ridda Ali
Abstract: Obesity has become a global public health issue. Based on the 2018 Health Survey for England (HSE), 67% of men and 60% of women were classified as overweight (Body Mass Index (BMI) of 25 to 29.9) or obese (BMI of 30 or more). HSE’s reports indicate that over half of adults (56%) were at higher risk of developing chronic illnesses (e.g. diabetes, stroke, arthritis, etc.) due to obesity.
Existing research in the area of weight management (WM) effectiveness is hindered due to attrition, adherence and abandonment. Hence, this study aims to: 1) identify the measures used in the MoreLife (provider of weight management programmes to individuals, families and communities) dataset and level of missing data for each variable; 2) discover if attendance at the start of the program can be predicted from psychometric and demographic information and to what level of reliability; 3) discover if the proportion of completion, amongst those who attend at the start the program, can be predicted from psychometric and demographic information and to what level of reliability.
Various methods of developing prediction models will be explored: all subsets regression, random forest and gradient boosting machine with cross validation. If necessary, multiple imputation will be explored for modest levels of missing or incomplete data.
If you wish to ask the presenters any questions or offer feedback, please email Michael Kettles LIDA Communications and Engagement Officer, who will pass these on. Thank you.