Eric Brill’s antique radio collection is taking over his living room. Twenty of the wood and vacuum tube, AM-only devices, some dating back to the 1930s, are crowded into Brill’s house.
It’s an impressive assortment for someone who had no interest in radios six months ago. Brill discovered them on eBay, where he’s the vice president of research and head of eBay’s research labs. Millions of other items compete for his attention on the site, including bestsellers and daily deals. What caught Brill was a particularly beautiful radio. He just stumbled upon it and became an avid collector.
Brill is now trying to make serendipity a regular occurrence on eBay. One of his labs’ significant projects is Discover, an alternate view of eBay’s inventory. Auction items you’d normally not see appear in a bunch of colorful tiles on a single screen or can be made into a slide show that advances with a single click or swipe on a tablet.
Products that make the cut on Discover may be a purse that looks like a Rubik’s Cube, a personalized Hogwarts’ acceptance letter, vintage World War II combat boots, a first edition of Charles Darwin’s The Descent of Man, a set of dog hoodies and a painting said to be from a “17th-century Flemish old master.”
What you won’t find on Discover is an iPad. “We already have enough places that surface those,” says Brill. “With Discover we’re saying, ‘We think you’ll be interested in this.’”
Sites such as eBay, Amazon and Netflix are always looking for new ways to surface what they believe you will want to click on. One of the most well-established methods for doing this is collaborative filtering—the analysis of user preferences and history—to make recommendations. The filtering is a mainstay at all three sites.
EBay’s serendipitous Discover goes further by serving up results that are a complete surprise to the user. To identify the most interesting goods on the site, the researchers on the Discover project wrote an algorithm that measures factors including how much people interact with a listing (clicking on it, returning to it, forwarding it to friends), seller history (the seller is known to offer interesting things) and textual analysis (a longer seller’s description of the product indicates more passion about the item).
The hope is to recreate the unexpected discoveries that are familiar to shoppers in brick-and-mortar stores. “When you go grocery shopping, you’ll have a list of 3 things and come home with 100,” notes Brill. “Serendipity is built into the layout of the store.”
EBay isn’t ready to say yet how, when or whether Discover will be incorporated into the regular site. The link to Discover is on the company’s testing ground site, buried at the bottom of eBay’s home page.
This is likely not to last. EBay is in the middle of a broader push to combine its data with faster, better software that make the site more inviting and useful. The San Jose company has bulked up its research staff with high-profile hires like Brill, who worked for Microsoft as a research manager, and Dennis Decoste, a machine learning expert who came from Facebook.
EBay Chief Executive John Donohoe vowed to revitalize the core Marketplaces business when he took over in 2008. Total volume of goods sold in dollars, known as gross merchandise volume or GMV (excluding vehicles), was up 14% for the U.S. in the June quarter, on par with the e-commerce industry. In the prior quarter, however, eBay’s 10% GMV increase lagged the industry’s 12% growth. EBay’s average selling price has risen as it has successfully steered more buyers to so-called trusted sellers who command premiums. Same-store sales for top-rated sellers in its largest markets (the U.S., U.K. and Germany) were up 18% in the first quarter and 22% in the second.
EBay doesn’t need more traffic. Its trouble has been converting visitors to buyers and getting them to buy more. Colin Gillis, a senior technology analyst at BGC Partners in New York, says Discover will help.
“EBay’s long tail was historically considered a weakness,” says analyst Gillis. But if Discover catches on, it could transform eBay’s massive collection of one-off, unusual items into an advantage against more mainstream competitors like Amazon, he says. “Data is what will take eBay above market-rate growth,” he contends.
EBay’s mess of data is both a gift and a complication. The company says it has more than 97 million active users worldwide, approximately 200 million live product listings and more than 10 million new and deleted items daily. Because eBay handles its own payment processing through subsidiary PayPal, the company can track each step of a sale. That information gets distilled into a comprehensive (though anonymous) dataset of eBay shopper behavior.
EBay labs employs economists, machine learning and machine vision experts, data miners, natural language processors and human computer interaction specialists. Neel Sundaresan, a senior director at the labs, says the economists are analyzing the incentive effects of free shipping, the profit impact of including more photos on sellers’ listings and the ability of reputable sellers to extract a premium from buyers.
Sundaresan says accessing and analyzing the company’s data used to be laborious. These days his staffers examine millions of rows of data at once and are urged to build their own products, similar to a startup. Dennis Decoste, who joined eBay as a research director in January, credits the shift to advances in computing hardware and algorithms. EBay also adopted new tools for big-data processing, such as the open-source software framework Hadoop and a system called Mobius that was developed in its labs. “Machine learning is in a golden age,” says Decoste. “We’re able to deal with all this data and see where it leads us.”
Earlier search improvements incubated in eBay’s labs have gone on to become standard features on the site. In 2008 it introduced a “best match” algorithm that displays the most relevant listings to what people have typed into the search box, even when the terms don’t obviously match. A recommendation system that suggests equivalent items to bid on if, say, you lose an auction, launched in 2009. Last year eBay released query suggestions, a feature that displays related searches and popular items below the search box as you type in your query.
The Discover experience can be addictive when each item leads to curious or exotic products. It can, however, disappoint by drawing attention to things like epoxy resin (too obscure) or cellphone chargers (not so delightful). The algorithm can be taught to make better choices via like and dislike buttons. Users can also pursue what Brill calls “constrained serendipity” by typing terms into a search box or picking categories through a drop-down menu.
At the same time Discover resists complete customization. Someone who is solely interested in baseball cards will not exclusively see baseball cards on Discover. “We would bias a bit towards that preference,” says Brill. “But we still want to surprise that person.” He says it’s too early to assess whether Discover boosts sales but points to user comments as evidence of satisfaction. One shopper sent in a note of thanks after he found the kind of bicycle he had as a child.
EBay doesn’t care if the sentiment is nostalgia or excitement, as long as it’s positive. Brill says he’s living proof. “No one could have studied me and said, ‘This is an antique radio lover,’” says Brill. “But the ability to see our inventory in this way is making me buy these things happily.”
[EDIT NOTE] This article is from the August 22, 2011 issue of Forbes magazine.
- machine learning
- vacuum tube
- combat boots
- Chief Executive