Never Miss a Beat: Get a snapshot of the issues affecting the IT industry straight to your inbox.
EBay has been gathering data and employing machine learning to connect more buyers to sellers and increase the trustworthiness of transactions. The use-cases are worth exploring for any IT professional looking to help improve a company's bottom line by applying machine learning to its customer-facing applications.
September 6, 2016
7 Min Read
(Image: eBay. The eBay Inc. logo is a trademark of eBay Inc. Used with permission.)
13 Ways Machine Learning Can Steer You Wrong
13 Ways Machine Learning Can Steer You Wrong (Click image for larger view and slideshow.)
eBay is applying the lessons of machine learning to improve the ways customers interact with the site. For four years, eBay has been collecting customer search data, along with search click-through rates and other customer interaction data, and feeding the information into its machine learning system.
In an interview with InformationWeek, Dan Fain, the company's VP of engineering, outlined the steps being taken, and the business motivation behind them. The use-cases are worth exploring for any IT professional looking to help improve a company's bottom line by applying machine learning to customer-facing applications.
Search, customer click patterns, language translations, item recommendations, and image analysis are among the key ways machine learning is being put to use at eBay, according to Fain.
[Want to see how eBay has already pioneered the use of big data? Read How eBay's Kylin Tool Makes Sense of Big Data.]
The first, and most important, application of machine learning is to improve search functions on eBay.com, according to Fain. "We have several machine learning models working behind the scenes to ensure we get the best search results," said Fain in our interview.
With more than 1 billion items for sale, many things can go wrong in a search on eBay. For example, much depends upon the words and phrasing chosen by the customer. For example, when someone searches "sewing machines," he or she probably wants a tool with which to do their own sewing.
But machine learning has shown that people are sometimes looking for a sewing machine their grandmother used to use, or even a "collectible" to fit into a set of machines with meaning to the collector.
The search triggers a check of eBay's inventory of sewing machine models for sale. The machine learning models may be able to gauge some intent from the history of the visitor doing the asking. They also check the behavior of other visitors searching for a sewing machine on the site. Did they browse and go away? Did they stop when they found a modern model? In what ways did antique sewing machine hunters display their interest?
Fain cited another search phrase, "Swiss watch with leather straps," which raises different challenges for the site that machine learning can address. A normal search would return a long list of Swiss watches and an equally long list of leather bands. If the person doing the search wants those two things as one product, such search results will not be very useful.
"In this query, one category should not exclude the other," said Fain. Even though such a search might occur only a few times a year, the system takes the long tail into consideration as it comes up with the best matches. From analysis of previous searches, the eBay search engine can connect several words that only infrequently occur together, Fain said.
The search engine has been schooled to carefully judge what category a search belongs in, then check the contents of the category against other clues about what interests the searcher. In the search example above, one clue would be an already brisk business in a popular type of Swiss watch on a leather band.
At the end of 2015, Fain said, eBay reported 162 million active buyers. By August 2016, the number had grown to 164 million. The modest increase is significant to a company with flat or declining revenues. The competition is fierce, and customers have other places to go for e-commerce.
The more Fain's team can apply machine learning to ensure customers find what they're searching for, the more likely a search will lead to a sale. It's also more likely prospective buyers will trust the information they're getting.
Machine learning also makes the site more effective at fielding queries coming from foreign buyers, who may be looking at a description of goods in their native language and not necessarily understand the conventions of the originator's language. For example, a search for Burberry handbags with a metal charm has to be handled differently in Spanish than it would be in English.
eBay's machine learning-based translations are making it easier than ever for site visitors who do not speak English to complete purchases. It's important for customers to be able to understand the presentation and value of what's being offered, regardless of where they are in the world or what language they speak.
To that end, Fain said eBay has added a "Best Match" algorithm to the search engine to analyze what's known about the buyer, what's popular among the items being sought, and how particular items might be of value to the prospective buyer. Such matching reflects eBay's "largest-scale application of machine learning," said Fain. "It is a powerful tool for surfacing deals."
For instance, in the sewing machine search example above, a Best Match response might spotlight machines based on value for price. If the searcher is a desktop user, eBay will have space on the screen to display the ranked Best Match results, along with additional categories the visitor might search, such as "antiques" or "collectibles."
On smartphones and other mobile devices there's not enough screen real estate to present much alternative information. That's a problem, because 50% of the business conducted on eBay now involves some form of mobile device, according to Fain. Mobile users get a link to the alternative categories on a new screen if they wish to take another step.
In addition to trying to make the search as relevant as possible, eBay has a big interest in making it reflect marketplace value. Results showing items at a good value for the price are displayed ahead of those with lesser value.
Trustworthiness of a transaction is also a high priority for eBay. Machine learning is applied to discern which factors indicate trustworthiness -- such as the rate of successful completions of sales by a seller -- and which do not. Conversely, if there's a string of complaints or issues with a seller, the result gets knocked down to a lower spot on the list.
By using machine learning, eBay is aiming to make sure both buyer and seller walk away satisfied. As a transaction takes shape, the system is asking, "How likely is it that this transaction will meet our standards?" Fain said. "The results are based on the strength of the evidence, determined by machine learning."
Looking Beyond the Click
Fain compared his five years working at Yahoo -- where the main evidence he had to work with was page clicks -- with the wealth of information available about eBay users.
Clicks tell the system that something has gotten a visitor's attention, but not much else, he said. At eBay, the system can study the click stream that leads to a purchase. "Buying is an incredibly powerful piece of evidence" to feed into a machine learning system, and eBay is in a strong position to collect purchasing evidence and make effective use of it.
Now, eBay engineering is using machine learning to better understand the quality of images used on eBay and to determine what makes effective image quality in a sale. It can employ the powerful parallel processing of graphical processing units in a cluster to analyze visitor interactions with images. "With the dedicated processing of GPUs, we can blast through way more examples," deriving lessons from large, real-world, eBay data sets, said Fain.
While he couldn't cite the number of servers dedicated to machine learning, Fain said they amounted to multiple clusters. EBay has found machine learning a valuable tool and is investing in more machine learning hardware, hiring technical people with machine learning skills, and launching additional machine learning projects. "Machine learning is a big investment for us," he said.
About the Author(s)
Editor at Large, Cloud
Charles Babcock is an editor-at-large for InformationWeek and author of Management Strategies for the Cloud Revolution, a McGraw-Hill book. He is the former editor-in-chief of Digital News, former software editor of Computerworld and former technology editor of Interactive Week. He is a graduate of Syracuse University where he obtained a bachelor's degree in journalism. He joined the publication in 2003.
You May Also Like
The ultimate survival guide to SOC 2 compliance
5 key areas for improved automation in InfoSec compliance
NIST Cybersecurity Framework 2.0: Changes, impacts, and opportunities for your InfoSec program
Edge Computing Bridges IT and OT People, Process, and Technology
How a trading floor continues its operations during COVID-19 lockdown