At eBay, machine learning technology has been breaking down borders for years. Starting in Russia, and now also in France, Italy, and Spain, eBay is leading the industry by applying automatic machine translation to commerce. When buyers search in these countries, eBay translates their request, and responds with relevant translated inventory from other countries in other languages. eBay is also pioneering powerful new ways to leverage machine learning technology to enhance the speed, accuracy and relevance of search on the eBay platform.
Translating search queries is an important step in providing eBay’s global customers with localized shopping experiences. However, in addition to translation, machine learning can produce striking efficiencies when applied to traditionally human cognitive functions such as perception and visual processing. According to some, machine learning could even alter how we think about search technology altogether. “As machines get better at decoding natural language, commerce should become increasingly conversational -- eventually rendering the search box redundant,” eBay CEO Devin Wenig recently wrote.
Selcuk Kopru is a Research Scientist working on integrating machine learning with search on eBay. “Machine learning is especially helpful at problems that require human cognitive capabilities like perception, language processing and visual processing,” he said. “At eBay, we work to anticipate and simulate these capabilities.”
Relevance and Prediction
As a few examples of how machine learning can enhance search and shopping experiences, Kopru points to predictive and item categorization tasks. “We apply machine learning techniques to item-to-product matching, price prediction and item categorization tasks on eBay,” Kopru said. “We also employ them for attribute extraction, generating the proper names of browse nodes, filtering product reviews and more. Machine learning helps us optimize the relevance of shoppers’ search and navigation experiences.”
Keywords are not enough for optimized search experiences, Kopru added. “Search has moved well beyond simple keyword matching,” he said. “We have seen that extracting semantics from item titles and descriptions using machine learning algorithms has helped to improve relevance in our search experiences.”
The largest scale application of machine learning technology at eBay is currently Best Match, the algorithm used to optimize relevance for buyers during their shopping experiences. Best Match analyzes everything from item popularity to potential value to the buyer, to terms of service such as return policies. It is a powerful tool for surfacing deals.
“Shoppers want relevance, but they also want great deals,” said eBay Technical Fellow David Goldberg. “Those deals might be auctions or Buy It Now items, and might come from an individual seller or a large retailer. Balancing all these things and finding the right match is a challenging machine learning problem that is unique to eBay.”
“We are also applying statistical learning to optimize the whole page: reacting to the users’ actions we can reorder and prioritize the content in the search page, providing search guidance or access to top deals and top products in appropriate context while minimizing distractions” said Alex Cozzi from the Search Science team.
At a Tipping Point
This is no ordinary time for machine learning, Kopru emphasized. “We are in the golden age of machine learning as applied to search,” he said. “Open sourced machine learning libraries enable applications for us without the need to implement everything from scratch. We can also run very complex and deep models with today’s computing clusters. These enabling technologies let us model new search experiences from many angles faster than we ever could before.”
Our CEO on Machines and Cognition
Machine translation, machine learning, search and artificial intelligence are all advancing in tandem at eBay. In a recent post, eBay CEO Devin Wenig had much to say about the promise of artificial intelligence and machine learning technologies patterned after human cognitive capabilities.
“The landscape is changing rapidly, particularly in the last six to eight months,” Wenig wrote. “We have seen radical sharpening of the intelligence of machines – computers can now use image recognition for diagnosing diseases or developing scientific theories.
“We already use machine learning algorithms to recognize objects in listings, find similar products, and rank recommendations,” Wenig added. “And we deploy AI in various areas, from structured data to machine translation to risk and fraud management. In the next few years, we’ll witness an unprecedented convergence of technology, commerce and consumer expectations.”