Making AI Human Again with Explainable AI

This year’s Annual Analytics Summit was again held at the IET, Savoy Place in London, with its magnificent views from the 3rd Floor Balcony stretching from St Paul’s in the east round past the London Eye to the Palace of Westminster in the west. The summit held by The OR Society in association with the Royal Statistical Society attracted a healthy number of delegates and exhibitors, as well as outstanding speakers. 

The first two presentations offered interesting perspectives on issues associated with the use of analytics and the application of AI today. Philip Pilgerstorfer, Data Scientist at QuantumBlack, opened with Making AI human again with explainable AI.

In some countries there is a legal requirement to disclose how, for example, a risk score for the insured/assessed had been arrived at. Individuals dissatisfied with the decisions made about their level of risk can request this information, so it has to be available and sufficiently transparent for them to understand the calculations. 

Explainable AI (XAI) can be used to meet the legal requirements for those cases in those industries. It can illustrate that no  discrimination actually took place when assessing risk; it can show that, for example, no gender bias had affected risk scoring for the insured or assessed. XAI too, can, when combined with predictive modelling, provide pointers to underlying reasons for making such risk score decisions and can help provide global understanding too.

XAI can be regarded as model agnostic insomuch as its principles can be applied to a wide number of models to provide “explainability” of how models are applied and introduced the concept of LIME (Locally Interpretable Model-agnostic Explanation).

Philip Pilgerstorfer finished his presentation with comments and slides in support of using a Shapley Additive Explanation Process (SHAP) to help illustrate explainability in AI, and how SHAP can be used to illustrate highly individualized explanations.