These are the most interesting articles on AI research that came out this year. It combines the latest artificial intelligence (AI) and data science advances. It is in order of when it happened, and there is a link to an article that goes into more depth.

FairFoody: Bringing in Fairness in Food Delivery

Along with the fast growth and rise to prominence of food delivery platforms, concerns have also grown about how the gig workers who make this growth possible are paid. Their analysis of data from a real-world food delivery platform in three big cities in India shows that delivery agents earn a lot of different amounts of money. In this paper, the researchers pose the problem of ensuring that agents get paid fairly, and that food is delivered on time.

The researchers show that the problem is not only NP-hard but can't be solved in a polynomial amount of time. However, they got around this problem by making a new algorithm for matching called FairFoody. Extensive experiments with real-world food delivery datasets show that FairFoody improves income distribution by up to 10 times compared to baseline strategies while having little effect on the customer experience.

Field Study in Deploying Restless Multi-Armed Bandits: Assisting Non-Profits in Improving Maternal and Child Health

The widespread use of cell phones has made it possible for non-profits to quickly get meaningful health information to their clients. This paper talks about their work to help non-profits that use automated messaging programs to send preventive care information to beneficiaries (new and expecting mothers) while pregnant and after giving birth. Unfortunately, a big problem with these kinds of programs is that many people who are supposed to get the information drop out.

Yet, non-profits often don't have enough health workers (or time) to make essential service calls to talk to beneficiaries live and keep them from losing interest. So researchers created a system called Restless Multi-Armed Bandits (RMABs) to help non-profits make the most of this limited resource. One of this system's most critical technical contributions is a new way to cluster offline historical data to determine unknown RMAB parameters.

Furthermore, the second big thing they do is evaluate their RMAB system with the help of an NGO through a study on how to improve service quality in the real world. Over seven weeks, the study compared ways to enhance service calls for 23003 participants to stop engagement drops. The researchers show that the RMAB group does better than other comparison groups in a statistically significant way, reducing engagement drops by about 30%. It is the first study to show how RMABs can be used to improve public health in the real world. The researchers are giving their RMAB system to the NGO to be used in the real world.

How Private Is Your RL Policy? An Inverse RL Based Analysis Framework

Reinforcement Learning (RL) enables agents to learn how to perform various tasks from scratch. In areas like self-driving cars, recommendation systems, and more, We could break privacy if the best RL policies learned to remember any part of the private reward. Researchers look at differentially-private RL policies from RL algorithms like Value Iteration, Deep Q Networks, and Vanilla Proximal Policy Optimization.

The researchers propose a new Privacy-Aware Inverse RL (PRIL) analysis framework that uses reward reconstruction as an adversarial attack on private policies that agents may use. They do this with the reward reconstruction attack, which uses an Inverse RL algorithm to determine the original reward from a policy that protects privacy. If the agent uses a tightly private approach, it will be hard for an adversary to determine the original reward function. Using this framework, the researchers try out different versions of the FrozenLake domain with varying levels of complexity to see how well the private algorithms protect privacy.

Based on the analysis, the researchers think there is a difference between the level of privacy currently available and the level of confidentiality needed to protect reward functions in RL. The researchers measured the distance between the original and reconstructed rewards. It lets them figure out how well each private policy covers the reward function.