Data-Driven Decisions Lab (3DL)
The abundance of data and improvements in computational performance have led to an unprecedented development of machine learning models in the past decade. However, the degree to which such models can affect real-world decision making still needs to be better understood. Led by Agni Orfanoudaki, Associate Professor of Operations Management, the Data-Driven Decisions Lab (3DL) aims to address this challenge and bridge the gap between predictions and prescriptions.
Our group conducts theoretical and empirical research employing and developing novel techniques in the fields of machine learning, optimisation and stochastic processes. Using datasets from industry and academic partners, and guided by conversations with practitioners, the members of 3DL focus on proposing new methodologies and quantitative frameworks in multiple application areas. These range from healthcare to revenue management and infrastructure.
There are currently four broad research streams within the group:
- Interpretable machine learning
- Dynamic decision making
- Healthcare analytics
- Algorithmic insurance
Interpretable machine learning
We aim to introduce a new set of machine learning tools that can bring about the adoption and integration of data-driven technologies into practice. To enable better human-algorithm engagement, 3DL members develop new transparent general-purpose machine learning algorithms that use tree-based architectures and mixed integer optimisation techniques.
We also study the interactions between algorithms and decision-makers, aiming to identify new hybrid models that incorporate human intuition into machine learning model training and deployment. The lab’s work highlights the importance of interpretability, providing an increasing amount of evidence that transparency in machine learning does not need to come at the expense of accuracy.
Analytics are poised to transform multiple areas of the healthcare industry. To this end, 3DL has established a series of collaborations with medical investigators across the world to create predictive and prescriptive models that can advance decision-making in healthcare. Our work is based on academic partnerships with healthcare practitioners and uses clinical data to improve the quality of care and operational outcomes for a wide range of conditions and medical departments.
In addition to clinical outcomes, the 3DL team delves into questions related to fairness and equity. Our goal is to minimise inequalities while improving the patient experience in healthcare systems with limited resources.
As machine learning algorithms start to get integrated into the decision-making process of companies and organisations, insurance products are being developed to protect their owners from liability risk. To propel the adoption of data-driven technologies, 3DL has set the foundations of algorithmic insurance.
This new class of insurance products transfers the risk of algorithm usage from decision-makers to specialist insurance providers, offering a financial vehicle that absorbs economic losses from algorithmic errors. The 3DL group focuses on the development of quantitative frameworks for pricing and risk estimation of algorithmic liability.
Dynamic decision making
Real-world decision making is highly dynamic and models that focus on a single time horizon tend to have difficulty predicting the implications of decisions. This risks long-lasting organisational performance in favour of short-term gains.
Members of 3DL try to develop new quantitative approaches that optimise decision-making in multi-period time horizons, identifying the policies that maximise long-term objectives. This work uses a combination of reinforcement learning and stochastic process techniques to improve the performance in areas where decision-making is sequential. This could include sectors like airline revenue management and healthcare.