CICS Team Wins ProjectX, International Machine Learning Research Competition
Content
A team of Manning College of Information and Computer Sciences (CICS) undergraduate student researchers, advised by doctoral student Cooper Sigrist, recently won the ProjectX Human-Computer Interaction prize at the University of Toronto for their project, “Recommendation Diversity Worth Caring About.”
The winning team focused their research on recommendation systems, algorithms that control the content users are exposed to while shopping, researching, or browsing social media. Recommendation systems are designed to help users find what they’re looking for faster and make suggestions for what to view next. However, as these systems collect data about users and the content they consume, the scope of the content they suggest becomes narrower—often resulting in presenting information and content that the user likely already agrees with or that targets their specific background and identity, risking the creation of echo chambers and deeper political and social divides.
“Our team was concerned about how dangerous poorly optimized recommender systems can be,” explains team lead and co-lead author Ryan Bahlous-Boldi ‘25. “With social media at the center of controversial and extreme political polarization, we thought recommender systems would be an especially important and relevant focus for our research.”
To combat echo chambers and homophily—recommendations for like content that targets people from similar backgrounds or identities—in recommendation systems, the team first proposes the introduction of two diversity metrics to a sample recommender algorithm. The first of these metrics uses a clustering approach on user embeddings, in which the group assigns numerical values to user interests to visualize whether users are moving closer to or away from an echo chamber. The second metric utilizes a set difference approach to determine how diverse the recommendations are, analyzing a set of recommendations to determine their diversity when compared to other items in the set and content the user has previously encountered, while maintaining the user’s minimum interest threshold.
Using these metrics, the group proposes an augmentation to a recommendation system that prioritizes diversity preservation and incorporates the use of a disaggregated form of preference prediction, which predicts user preference for certain item features rather than the item itself. Finally, the team utilizes a form of selection known as Lexicase selection that yields content recommendations that are diverse enough to move users further away from echo chambers while matching user preferences for item features. Using these metrics allows the team to introduce diversity granularly, further reducing the likelihood of forming instances of echo chambers and homophily.
“I leaned on some of the knowledge I’ve gained from previous research experiences in evolutionary computation regarding Lexicase selection,” explains Bahlous-Boldi. "The idea clicked for our team that recommender systems could benefit heavily from this kind of diversity, and as our research suggests, they do.”
ProjectX is a five-month machine learning research competition hosted by the University of Toronto Artificial Intelligence Group, in which undergraduate teams led by graduate or faculty advisors produce original research for a chance to win part of a $40,000 CAD prize pool. The 2022 competition focused on Human-Computer and Human-Human Interactions and included fifteen teams from Canada, the United States, and the United Kingdom.
The undergraduate research team—including Bahlous-Boldi, co-lead author Aadam Lokhandwala, Alexander Lavrenenko, Yuval Shechter, and Edward Annatone, now a master’s student at CICS—credits Sigrist and the collaborative environment fostered by the conference for their success.
“Cooper was an amazing guide for us. He was hands-on enough to ensure we didn’t stray too far but gave us the freedom to explore our ideas and determine the direction we wanted to take our research. He did not hold back when there were issues with the project and helped us immensely when high-level research skills were required,” says Bahlous-Boldi. “There were also times that I could feel our team suffering from imposter syndrome,” he continues. “He had faith in us and gave us the confidence we needed when the going got tough.”
“I’ve been advising undergrads for the past three years, and this felt like the culmination of that work,” says Sigrist. “Undergraduate students have a lot of passion and are creative as hell, and with someone to help them focus that passion, they can do wonderful work.”