A Blog by Jonathan Low


Jul 10, 2019

How Algorithms At Zappos Are Learning To Teach Themselves

The goal is to make it easier for customers to find what they seek by reducing the number of words they have to use or the number of times they have to click.

Self-learning algorithms use the principles of natural selection to find the optimal solution, which is then absorbed by the next generation of models. JL

Jared Council reports in the Wall Street Journal:

Genetic algorithms produce solutions to a problem using natural selection, including reproduction and mutation, (to) produce the optimal or “fittest” solution. Self-learning algorithms eliminate search mishaps and increase the relevancy of its search engine. The algorithm that performs best on a “relevance test” has the greatest chance of its traits passed to the next generation. The best-performing algorithm is placed live on the website until another performs better. “We’ve seen a decrease in repeat searches, people trying to add more words to the search and fewer clicks to find the right results. The results imply "it is easier to find stuff."
Zappos LLC, an online seller of shoes and apparel, said a self-learning algorithm has shown promise in solving one of its most perplexing search-engine issues: irrelevant results.
The 20-year-old Las Vegas-based company, which Amazon.com Inc. purchased in 2009, relies on its internal search engine to help customers find what they’re looking for. But for years, the search engine would get tripped up by a variety of words and phrases. For instance, in past searches for brand names that include a color—such as “Red Wing boots”—the search engine would see the color and deliver results with that color—in this case, red boots. Searches for “dress shirt” would include dress shirts and dresses.
Ameen Kazerouni, the company’s lead data scientist, said his team about two years ago started testing a decades-old technique known as a genetic algorithm as a potential solution. The self-learning algorithm started eliminating such search mishaps within the first year, he said, and has become a key factor in increasing the overall relevancy of its search engine.
“We’ve seen a decrease in repeated searches—people trying to add more and more words to the search term...and fewer clicks in what it takes to find the right results,” Mr. Kazerouni said. The results imply “that it is easier to find stuff—which is something we’ve not been able to achieve before,” he said.
Genetic algorithms produce various solutions to a problem and use principles of natural selection, including reproduction and mutation, with the goal of producing the optimal or “fittest” solution. For instance, genetic algorithms can be used to help logistics companies discover the best routes for delivery drivers.
While genetic algorithms have been around since the 1970s, advances in computing power and speed have increased their appeal.

“For a long time, speed was the reason a lot of people didn’t use this stuff. Now as computers are starting to speed up…people are starting to use techniques that beforehand were just too slow,” said Eli Finkelshteyn, chief executive of Constructor.io, which also uses genetic algorithms. The company sells search and product-discovery software that powers e-commerce websites.
Unlike with traditional search engines, Mr. Finkelshteyn said, those utilizing genetic algorithms in particular and machine-learning algorithms in general can integrate user behavior after a search as feedback. The system can use that feedback to learn what’s relevant for future searches, which can ultimately boost sales.
In 2017, Zappos saw roughly a million unique terms searched on its site each month. Its search engine had to match those terms with products in its 100,000-plus item catalog. Tackling search-result mistakes—such as showing shorts when people search for “classic short,” a popular Ugg boot style—seemed to be a complex math problem, Mr. Kazerouni said. His team thought genetic algorithms, which have been used for complex math problems in logistics, telecom and other industries, could help.
The Zappos team built its system in house and designed it to produce algorithms that could parse out the intent of a search phrase. One algorithm might see the word “dress” as a strong signal for fetching dresses, but another might play down that signal and pay more attention to the surrounding words.
In that scenario, the algorithm that performs the best on an internal “relevance test” developed by Zappos—which simulates how users engage with search results—would have the greatest chance of seeing its traits passed to the next generation. The best-performing algorithm is placed live on the website until another one performs better. The company wouldn’t disclose metrics around its search engine’s performance, but it said the relevance scores for its genetic algorithm initially were as low as 0.0001 on a scale of zero to 1 and as high as 0.8 in the latest generation.
The Zappos team later developed two other genetic algorithm engines. The company uses all three in tandem to come up with better search results.
“If a person wants to buy some shoes, and they include the word ‘dress’ and the site starts giving them dresses instead of dress shoes, they’re going to get fed up and go to a competitor,” Mr. Finkelshteyn said. “It’s incredibly important to give users relevant and appealing results so they stay on site and keep looking.”


Post a Comment