MSc Student project suggestions and notes
Updated: 12/12/2021
I am currently adopting MSc students who will be taking my Semester 2 module ECS7005P "Risk assessment for data science and AI". On this module students will learn how to use specialist software to build causal models (Bayesian networks) that enable risk assessment and decision support for a broad class of problems where there are insufficient data for pure machine learning algorithms to provide useful or accurate results. We call this the 'smart data' approach - in contrast to the 'big data' approach.This 5 minute video explains why smart data, rather than big data, can achieve 'true AI": https://youtu.be/-7vSiWRasxY
For example, imagine you were trying to learn from data whether customers applying for loans are likely to default; the data available is about people previously given loans and whether or not they defaulted- but we have no data about people who were refused loans. Or imagine trying to learn from data whether patients with particular risk factors are likely to suffer from a particular disease; the data available are about previous patients known to have suffered from the disease, but we have no data about people who did not suffer from the disease and no data about the kind of interventions or treatments any of the patients had. In cases such as these we must use knowledge to build an underlying causal model and then a combination of data and knowledge to 'paramaterize' the model.
Each project will use a combination of data and knowledge to develop a Bayesian network model for a working decision support system for one of the following kind of problems (although I am open to other well motivated suggestions):
- monitoring and diagnosis of a chronic medical condition
- early detection and monitoring of risk (for example: will students drop out of a course, are children likely to be harmed at home, will subscribers in a particular service renew at end of contract, etc)
- determining whether there is bias in a system that makes decisions about people (for example, automated systems that review job applicants, predictive policing systems etc.)
- determining whether a film will be a box office hit (purely data driven regression models do not make accurate predictions - see this paper -, so the objective is to incorporate knowledge in addition to the data in websites such as Themoviedb.org and imdb.com
- determining whether the evidence in a criminal investigation is likely to result in a successful prosecution
https://youtu.be/3KGYuLFMRSY
Two short 2-page overview articles of the smart data approach are:
- http://www.eecs.qmul.ac.uk/~norman/papers/BAYES-KNOWLEDGE_EUR15_H_Res.pdf
- http://www.eecs.qmul.ac.uk/~norman/papers/OAG_published_high_res.pdf
Previous highly successful projects:
- Prodhan, G (2020). Extending the range of COVID-19 risk factors in a Bayesian network model for personalised risk assessment. https://doi.org/https://doi.org/10.1101/2020.10.20.20215814
- Butcher, R.. (2020). Extending the range of symptoms in a Bayesian Network for the Predictive Diagnosis of COVID-19. https://doi.org/10.1101/2020.10.22.20217554
- Collins, R.. (2020). Bayesian network modelling for early diagnosis and prediction of Endometriosis. MedRxiv, 2020.11.04.20225946. https://doi.org/10.1101/2020.11.04.20225946
- Preekan, S (2020) Quantifying Reputational Risk with a Bayesian Network MedRxiv, 2020.11.04.20225946. https://doi.org/10.1101/2020.11.04.20225946
- Hoyal, S (2020) An integrated Bayesian Network providing decision support to front line workers in health and social care when children present injuries
- Orishko, O (2020) Abstractive Text Summarization for Book Reviews and Review Filtering in a Mobile Application