Prediction algorithms in AI use machine learning and statistics to make predictions about an event, given what we know now. Examples include whether a covid-19 patient will require ventilation, or whether a person seeking insurance will make a claim. These predictions can be used for planning and decision making. However, a limitation of these approaches is they cannot (and should not) be used to ask ‘what if’ questions. For example: ‘what if we give a patient CPAP rather than ventilation?’, ‘what if the person seeking insurance had a different ethnicity?’. The first example is important for decision making, while the second has implications for fairness (since it would be considered discriminatory to charge a different insurance premium based on ethnicity). In both cases, causal inference can help to enrich these prediction algorithms with ‘what if’ capabilities.
This event will present the outcomes of a Turing Institute ‘Theory and Methods Challenge Fortnights in Data Science and AI’, held in Manchester in February 2020. A challenge team, comprising twelve academics in predictive modelling, machine learning, and causal inference, will share the findings from the challenge – including perspectives on ‘counterfactual prediction’ and pilot work to address some of the methodological challenges involved. This is now topical work, as such methods can be used for decision support in the covid-19 pandemic.
Add comment