Arpit Agarwal combines theoretical computer science, machine learning, and artificial intelligence with information elicitation, preference/choice modeling, and economics to explore how algorithms and people interact. He focuses on online recommendation systems—how people choose items, how recommendations affect those choices, and how we may learn from this behavior to make better recommendations—to improve online interactions and optimize for high-value relationship building and long-term rewards.

“Many algorithms are designed to optimize for short-term reward, for people to purchase impulsive, low value items, which doesn’t benefit the consumer and can cause them to turn away from the company,” said Agarwal, who is a postdoctoral research fellow at Columbia University’s Data Science Institute (DSI). “We should be creating a win-win situation. Consumer behavior and algorithms can work harmoniously together to maximize long-term rewards for the consumer.”

Agarwal completed his doctorate in computer and information science at the University of Pennsylvania and a master’s degree in computer science and engineering at the Indian Institute of Science. He values the freedom and flexibility of his postdoctoral engagement at Columbia with DSI. “I have a lot of independence in my research trajectory. The relationships are not territorial or exclusive, and I have been able to collaborate with multiple people simultaneously,” he said.

For example, Agarwal works with Yashodhan Kanoria, an associate professor in the decision, risk, and operations division at Columbia Business School and a DSI affiliate, to develop models of human/algorithmic interactions that incorporate causal reasoning where the algorithm is aware of how its recommendations influence user options.

Agarwal also collaborates with computer science professor and DSI member Tim Roughgarden and doctoral student William Brown to design recommendations that not only maximize reward, but also continually show the user a diverse set of items.

Currently, models of human behavior, decision making processes, and utility functions are not updated based on interactions and do not include causal reasoning, according to Agarwal. “I am working on improving these models so that they respond to input from human engagement and choices. This is challenging because it requires more computational power. It can slow down the algorithm, limit capacity, and negatively affect scalability and growth.“

Incorporating feedback loops and diverse recommendations may be difficult, but Agarwal considers this work necessary for better consumer experiences and for better online decision making overall.

— Karina Alexanyan, Ph.D.