Ever wonder how Netflix knows what movies you like or how Pandora can turn you on to a band you’ve never heard of? It’s all based on recommendation algorithms: computer codes that analyze what you watch, listen to or buy and recommend something new that you might enjoy.
Thanks to a new grant from the National Science Foundation, Ithaca College Associate Professor of Computer Science Doug Turnbull is working to make those algorithms better and cheaper to make, so that more people can take advantage of the technology.
The grant, which totals $500,000, will be split between Turnbull and researchers at Cornell University. Turnbull and his students will work on their music app MegsRadio.fm, which uses a recommendation algorithm to suggest local musicians to listeners. The data provided from the app will be used by researchers at Cornell to test more effective algorithms.
Risky business
Testing changes to an algorithm can be a risky and expensive proposition, according to Turnbull. “Recommender systems are slow to evolve because every tweak to the system involves getting new users to try it out, and there’s a risk there,” he says.
For example, Amazon uses an algorithm to recommend items for shoppers to purchase. When it tests changes to its algorithm, the company could lose money if those changes result in lower sales. But if Amazon could test its algorithm against already existing data, it could reduce that risk.
It’s called “counterfactual learning,” and that’s what Turnbull and his colleagues at Cornell are studying. In short, counterfactual learning tests algorithms using already existing data on previous users’ choices. That way, the algorithm isn’t tested on actual live users.
But that process isn’t easy, since the existing data can’t match up perfectly with the new algorithm. Under the grant, researchers are working on ways to align the two, so that the data bears accurate results when used to test new algorithms.
It’s All About the Data
Studying counterfactual learning requires existing data from recommender algorithms. However, that data is valuable and companies like Amazon or Google won’t just hand it over to anyone who asks. That’s where MegsRadio comes in.
“Our role in this is to build the MegsRadio system to generate the kinds of data that we would need for this kind of learning and to interact and try to improve the playlist algorithms,” said Turnbull.
Turnbull and his students built MegsRadio using a prior grant from the NSF. It officially launched this past September. The app functions like Pandora, except that it recommends music from artists from in and around Ithaca that users might enjoy. It does that using a music recommendation algorithm that analyzes music waveforms, social data, listening history and user preferences.
The data produced by MegsRadio is sent to Cornell, where researchers use it to experiment with counterfactual learning and test tweaks to the recommender algorithm. They then send the new algorithm back to Turnbull to improve the quality of the listening experience on MegsRadio.
In the course of that back-and-forth between IC and Cornell, researchers hope to enhance the quality of counterfactual-based testing. That could pay dividends for existing and new companies interested in recommender algorithms, as they could test changes without the associated risks and costs.
Users, of course, stand to benefit in the form of more useful search engine results and better movie and music recommendations.
Turnbull also points towards economic benefits for the City of Ithaca. MegsRadio not only recommends local artists, it also tells users when and where those artists are performing. As MegsRadio gains users it could help drive people to bars and music venues, supporting local business owners and bolstering the local economy.
Experiential Learning
For Ithaca College students studying computer science, the MegsRadio project has been invaluable.
A team of 10 students are involved in developing the app and doing outreach in the community. Turnbull acts as a manager, letting students do the bulk of the work.
“In terms of how we approach and solve problems, it’s student-driven,” said Luke Waldner ’17, a computer science major who works on the project. “That’s really valuable especially for getting jobs in the industry because we have actual experience working with a team and doing actual problem-solving and self-motivated research.”
Mariah Flaim ’16, who helped build the MegsRadio iOS app as a student, says her experience helped her land job interviews after graduation. She now works as a technical consultant at the software company Red Hat. Flaim, like several other graduates, still helps work on MegsRadio.
“I love the team and I am passionate about the project,” she says. “I really believe that the apps will go places.”