Strategy Proof Machine Learning

Key Research Objectives
- Typical Machine Learning models assume that data is sourced cleanly from a distribution of interest. In reality, this is not case. In the real world, data is often provided by entities who have a vested interest in the decisions made by a Machine Learning model and thus may be motivated to submit data strategically in order to influence the decisions made by a Machine Learning algorithm.
- I aim to establish Machine Learning techniques which can work in more realistic settings in which data providers have their own motivations.
- Challenges in doing this include:
Providing an adequate modelling framework which captures the problem of learning from data providers who may have different incentives and capabilities.
Designing efficient optimisation techniques which elegantly deal with the optimisation problems which arise as a result of this modelling framework.
Effectively combining techniques from Mechanism Design with algorithms from Machine Learning.
- In addition, I aim to address other substantial differences between the standard theoretical Machine Learning setting and the practical environments in which Machine Learning is often deployed. For example, problems often present in practical settings include:
Feedback about decisions may be delayed.
The entities providing data (and their interests) may change over time.
The actions of one data provider may influence another to submit different data.
The set of decisions available to make may dynamically change over time.
- To tackle these kind of issues, I aim to extend existing models for Online Learning and Bandit Feedback.

Novelty and Methodologies used
- Standard models of manipulation in Machine Learning predominantly consider all powerful data providers who aim to cause as much harm as possible. We aim to consider data providers which are more likely to occur in reality - those with limited capabilities and their own objectives.
- Unfortunately this typically results in optimisation problems which are tremendously difficult to solve. As a result, we aim to derive new algorithms which can deal with this specific class of problems more efficiently than standard out of the box algorithms.
- We hope to use and apply techniques from Mechanism Design to aid us in updating existing Machine Learning algorithms. Our work is therefore at the intersection of Mechanism Design and Machine Learning, which is a relatively unexplored area of research.
In addition, the shift to considering problems which occur in practical Machine Learning settings, such as the ones above, naturally leads us to consider Online Learning and Bandit Feedback models for learning. Much of our work will consist of extending these approaches to address new practical considerations.