Algorithms Play Matchmaker to Fight 7.7% U.S. Joblessness
When Monique Nyampong graduated from Long Island University last May, she wished she had a headhunter who knew of employers with openings that would be right for her.
She got the help she needed in the form of a patented computer algorithm developed by AfterCollege Inc., an online job board that analyzed 12 years’ worth of data it has stored about its users and business clients to find the best positions for her. A month after registering her profile on the service, she found a management-training opportunity with Consolidated Edison Inc. sitting in her inbox. She’ll start the job in June.
“I felt like it gave me specifically what I wanted, like it was catered to me,” Nyampong, 25, who lives in Queens, New York, said of the service.
Job-search services like AfterCollege are racing to develop software that thinks and acts like a human recruiter, examining what kinds of workers and companies are drawn to each other so a computer can recommend vacancies to job seekers and candidates to hiring managers. With unemployment at 7.7 percent, more than 12 million Americans are looking for work while almost 4 million openings go unfilled, according to data from the Bureau of Labor Statistics. Targeted matches may help the labor market work better, according to Alvin Roth, an economist at Stanford University near Palo Alto, California.
“There are all kinds of inefficiencies that these firms are trying to solve,” said Roth, who shared last year’s Nobel Prize in economics for his research in matching markets. “It might be hard for me to know about the job that’s available. There might be some skill you’re looking for but you might not find me.”
San Francisco-based AfterCollege and competitors such as CareerBuilder Inc. and Burning Glass are taking advantage of mountains of information they have stored over the years. Using resumes that job-seekers registered, the services extracted data like the degrees and certifications workers had acquired, their geographic locations, past job titles and previous employers and when and how long they held positions. From the job descriptions employers posted, the services also collected and organized information like job title, employer, education and experience requirements of positions as well as the benefits they list.
Engineers then looked for patterns. What kind of worker tends to view, apply to and eventually land what kind of jobs? What kind of employer typically targets what kind of candidate?
That analysis forms the basis of the computerized recommendations. When a job-seeker like Nyampong tells AfterCollege that she graduated in May 2012 with a joint bachelor’s and master’s in accounting from Long Island University, and the website tracks her clicking into job advertisements A, B and C, AfterCollege’s computers use their memory of job leads that proved to be good for people like her in the past to find similar positions for her now.
“A traditional headhunter will connect people with jobs, but we’re using data to do it automatically,” said Roberto Angulo, chief executive officer of AfterCollege. “We know some things about people based on past behavior, and so we can do a good job predicting what are the good opportunities for you to check out.”
While preliminary results from the companies suggest their early attempts with recommendations are working, it’s not the first time technology held the promise of a better-functioning labor market. Once the Internet allowed companies to target more potential workers, and job-seekers to discover more openings, economists wondered if the jobless would find work more quickly and permanently bring down unemployment.
Hiring instead “turned into a needle-in-the-haystack problem, where people are besieged with tens of thousands of applications,” said Autor, who specializes in labor market issues.
The challenge, said Stanford’s Roth, was a lot like the early fumbles of online dating.
With little refereeing, men would flood the inboxes of attractive women, who in turn would ignore most of the messages they received. Men would then contact more women, which meant they were sending less personalized notes.
In dating as well as job-searching, “the interactions become less informative and less rewarding,” said Roth, who has designed systems that matched new doctors with hospitals and organ donors with patients.
To try to improve matches between its business clients and freelancers, oDesk Corp., the U.S.’s largest online market connecting businesses to remote contractors, started an experiment in 2011. The Redwood City, California-based company built an algorithm that suggested about six best-match candidates to a randomly selected pool of businesses, basing the formula on data like the freelancers’ feedback score from previous employers as well as how recently they finished their last project.
The suggestions worked: businesses receiving the recommendations filled almost 17 percent more of their technical vacancies, compared with employers that weren’t provided with the guidance. oDesk, which launched its online marketplace in 2005, now provides direct recommendations of freelancers to all of its clients.
“There’s no off-the-shelf answer” for computerized matchmaking in the job market, said John Horton, an economist with oDesk who documented the study. “That’s why oDesk has hired a big data-science team. We have to create the technology ourselves.”
Beyond the online freelancing market, jobless workers receiving government help finding employment are benefiting from the data-driven approach as well. Boston-based Burning Glass, which provides technology for public job-matching programs in states including New York, New Jersey and Oklahoma, coded 60 million people’s resumes to teach its software how people transitioned from one job to the next.
That allows the computer to recommend likely -- and sometimes unexpected -- next steps for the unemployed, said Chief Executive Officer Matt Sigelman. Some 55 percent of users said they received good matches for positions or occupations that they hadn’t considered, according to a survey conducted by the New Jersey’s Department of Labor and Workforce Development.
“Our way of matching is looking at what proves out in the job market,” said Sigelman. “It’s big data applied to the job market.”
The focus on generating recommendations is also gaining traction at Chicago-based CareerBuilder.com, which attracts more than 24 million visitors every month. Some 60 percent of the openings the website’s users now apply for are found through those suggestions, instead of keyword searches that users themselves perform, according to Chief Development Officer Hope Gurion. That figure was 45 percent at the beginning of 2010.
Still, important signals of a candidate’s employability, like dedication or enthusiasm, may be too human to quantify. More worrisome has been a tendency for companies to shrink their recruiting teams as they place too much hope on automating the screening process, according to Peter Cappelli, a labor economist and director of the Center for Human Resources at the University of Pennsylvania’s Wharton School in Philadelphia.
“The problem is that the match between organizations and people is tricky,” said Cappelli, author of “Why Good People Can’t Get Jobs.” “Trying to use software to do this is a little beyond the reach” of existing technology.
Those concerns haven’t stopped AfterCollege’s Angulo from banking on the kind of artificial intelligence that Amazon Inc. used to transform e-commerce and Pandora Media Inc. (P) applied to popularize Internet radio. Angulo, who co-founded the company in 1999 as an online job board for students and recent graduates, was so confident in his improved algorithm’s powers that last year he got rid of the search box that dominated AfterCollege’s homepage.
Job-seekers like Nyampong now fill out a profile of themselves instead, trusting the computer to do a better job than their own searches using keywords.
“I wanted to be open-minded,” said Nyampong, an accounting major, who spurned the traditional career path of working for a large international auditor. After registering with the service, “I just waited for the e-mail alerts” containing the job suggestions. “I applied to Con Edison as soon as I saw it.”
To contact the editor responsible for this story: Chris Wellisz at email@example.com