If you shop on Amazon, watch TV and movies through a streaming service, or keep up with current events through social media, chances are recommendation systems have shaped your online habits and preferences. These systems rely on algorithms to pull together a lot of data to arrive at a set of “top” choices for you.
Farzad Farnoud, assistant professor with the University of Virginia’s School of Engineering, wants to improve how algorithms are developed and tested to make the final “top” results more representative of individual choices and preferences. Farnoud holds joint appointments in the Charles L. Brown Department of Electrical and Computer Engineering and the Computer Science Department.
“I’ve always been motivated by problems that bridge mathematical models and practice—to establish fundamental limits while having a practical, and positive, effect,” Farnoud said. Farnoud’s research focuses on a type of algorithm that ranks a set of alternatives based on individual preferences.
Conventionally, this is performed using numerical data, e.g., scores on a scale from one to 10, in which one is worst and 10 is best. The problem is that each of us has our own subjective understanding of what each number in such a scale means: is seven out of 10 great, good or average? The interpretation varies from person to person and even for an individual at different times.
This subjectivity poses a challenge for algorithms designed to evaluate relative preferences within voting and other types of governance systems. A more robust approach collects individual preferences in the form of rankings and comparisons. Rankings and comparisons are more reliable because statements such as “A is better than B” do not require a scale. But rankings and comparisons come with their own computational challenges, driving development of scalable and efficient algorithms for large heterogeneous datasets.
In July 2019 the National Science Foundation’s computer information science and engineering program awarded Farnoud and Quanquan Gu $500,000 to develop decision-support algorithms that perform well with heterogeneous data. Gu is an assistant professor of computer science at the University of California, Los Angeles. The pair has a jump-start on this research, using preliminary results they produced with support from a UVA Engineering innovation award.
In addition to addressing algorithms’ reliability, Farnoud and Gu will explore how to improve algorithms’ performance in an interactive decision-making environment such as crowdsourcing. “Think of the guessing game 20 questions. You’re given a category – animal, vegetable or mineral. The player who asks the best questions to narrow down the possibilities the quickest will win. Similarly, our goal is to arrive at a decision by asking the most informative questions for the least cost,” Farnoud explained.
The grant fits well with Farnoud’s broader interests in information theory—the study of measures of information, its storage and communication. Information theory finds the boundaries of any algorithm’s accuracy by focusing on the quality of information and its relevance to a problem.