Starting Tuesday, the shopping discovery site Polyvore expands its expertise from just fashion to home goods, too. But this deceptively simple announcement reveals nothing about the technical hoops Polyvore’s developers have had to jump through to make this happen.
Polyvore is best known as a place where users can make collages, or “sets,” out of their fashionable finds. The 20-million user community certainly isn’t the only fashion discovery site out there, joined by the likes of Wanelo, Wish, Fancy and others.
But while the e-commerce sphere has clearly nailed down fashion, home decor has proved to be a difficult beast to tame. Pinterest, while not self-defined as a shopping site, has shown just how popular home goods curation can be. Why do competitors like Wanelo only dabble in home decorations and why is Polyvore only just committing to it now?
“It turns out home is a lot more complicated than fashion,” said Jess Lee, Polyvore CEO.
According to Lee, Polyvore’s had to revamp its search algorithms, which previously were fitted only for fashion. For example, the algorithms realized that there is only one use for a clothing item like a shirt—it goes on your body. But a lamp? That could go in a bedroom, bathroom, any room in the house. And that’s to say nothing of home items’ collisions with fashion.
Same Search, Different Meanings
“Because Polyvore supports both fashion & home, we often have to disambiguate queries and usage. For example, if you search for ‘glasses,’ do you mean drinking glasses or eyeglasses?” Or if you search for ‘floral,’ did you want floral dresses or floral throw pillows?” said Lee.
“We had to build separate search indexes for fashion and home, and then try to figure out which one you meant.”
Originally, when Polyvore began in 2007, founder Pasha Sadri used it to create mood boards while remodeling his home. But while users have technically been able to work with home goods since the beginning, the site’s search just wasn’t robust in that department. Polyvore’s search could only tell if you were looking at a lamp, and suggest other lamps. Starting today, it can categorize and add recommendations for stylistically similar lamps by popularity and taste.
“Training our machine learning classifiers and categorizers took more work because there was more ground to cover,” said Lee. “Generating product recommendations required a deep understanding of the product features that matter. For example, chevron prints on pillows and modern stylings for beds.”
Polyvore gets points for ambition, but don’t expect the service to be perfect. Unlike fashion, which shopping discovery sites have had years to perfect, home is just gaining traction. The more data Polyvore algorithms learn from users, the more accurate it will eventually become.
“Data quality is one of those areas where it’s impossible to achieve perfection and you can only asymptotically get closer,” said Lee. “Users always notice if results are bad, however they won’t applaud you for ‘not’ screwing up your search results and recommendations because it just feels like part of a natural and delightful user experience.”