Machine Learning for Recruiting at Amazon – Challenges and Opportunities

Amazon’s challenges and opportunities using machine learning for recruiting

Machine learning has been around for sixty years, but just recently a confluence of factors came together that have enabled the field to make exponential progress: enormously increased data, significantly improved algorithms, and substantially more powerful computer hardware.[1] As a result, machine learning is becoming ubiquitous. For example, in a recent study of 170 industrial organizations, 96% of respondents agreed or strongly agreed that machine learning is automating process-change management inside their organization. Two advantages of machine learning are self-adapting processes that can enhance customization, and self-repairing processes that can solve their own problems.[2] While machine learning cannot be used to draw causal conclusions, it can be deployed to make predictions such as personalized recommendations for customers, forecasting long-term customer loyalty, anticipating the future performance of employees, and rating the credit risk of loan applicants.[3]

As machine learning becomes more widely used, it is being applied to a growing number of industries and corporate functions. Recruiting is no exception. Today, companies take an average of 42 days at an average cost of $4,129 to fill each requisition. In addition, hiring mistakes are extremely costly: one study found that 41% of employers estimated a single bad hire costs at least $25,000, and 25% put the figure at $50,000 or more. When stakes are this high and the war for talent only becoming more competitive, companies must seek any edge.

Amazon was unusually well suited to enter the machine learning for recruitment space because it had already effectively deployed machine learning elsewhere in the company. Amazon pushed its machine learning capability by creating Echo and the Alexa voice platform that powers it, which was in effect Amazon’s Watson – a moonshot project that built a capability that could be leveraged throughout the company. Alexa “spurred a larger AI renaissance at the company,” enabling Amazon to apply the capabilities developed through Alexa to other products such as Fire TV, voice shopping, the Dash wand for Amazon fresh, and ultimately Amazon Web Services.[4] Machine learning is now embedded in Amazon’s culture.

With these important successes under its belt, Amazon recently attempted to integrate machine learning into its recruitment process. Amazon’s “experimental hiring tool used artificial intelligence to give job candidates scores ranging from one to five stars – much like shoppers rate products on Amazon.” Initially, Amazon had high hopes for the project: “Everyone wanted this holy grail…they literally wanted it to be an engine where I’m going to give you 100 resumes, it will spit out the top five, and we’ll hire those.” Unfortunately, the project stalled and was ultimately abandoned once it was determined that the machine-developed algorithm was biased against women: it penalized resumes that include the word “women’s,” and it favored candidates who described themselves using terms more commonly found on male engineers’ resumes such as “executed” and “captured.”[5] According to recruiting expert Kaya Payseno, Amazon made three critical mistakes: thinking the bias is coming from the machine, limiting the data set, and deriving future predictions from past events.[6] If Amazon had acknowledged these shortcomings, perhaps the project would still exist today.

Despite this setback, opportunities still exist for Amazon to learn from other organizations that are effectively leveraging machine learning within their recruitment process. According to Susan Poser and Sharad Sinha of Oracle, “with the right strategy, computers can find correlations that humans overlook, leading to better candidates.”[7] Companies are “training machine learning algorithms to help employees automate repetitive aspects of the recruitment process such as resume and application review;” they are also integrating machine learning into other aspects of the process such as talent sourcing and candidate screening and engagement. Indigo, Canada’s largest bookstore chain with 6,500 employees, used machine learning to reduce cost per hire by 71% and triple the quantity of qualified candidates in its applicant pool. Other technologies are helping companies tap into the large pool of passive candidates: for example, Entelo claims that its More Likely to Move algorithm can identify individuals who have a 30% likelihood of changing jobs within the next ninety days.[8] All of these are important opportunities for Amazon to consider as it builds its massive workforce.

When we look toward Amazon’s future with machine learning in its recruitment process, open questions remain including (1) what conditions are necessary for machine learning to add value not only to recruitment but also selection; (2) is machine bias equivalent to human bias; and (3) what can be done to eliminate machine bias in the future.

[1]Erik Brynjolfsson and Andrew McAfee. “What’s Driving the Machine Learning Explosion?” Harvard Business Review. July 18, 2017. https://hbr.org/2017/07/whats-driving-the-machine-learning-explosion

[2] Allan Alter, Sharad Sachdev, and H. James Wilson. “Business Processes are Learning to Hack Themselves.” Harvard Business Review. June 27, 2016. https://hbr.org/2016/06/business-processes-are-learning-to-hack-themselves

[3] Mike Yeomans. “What Every Manager Should Know About Machine Learning.” Harvard Business Review. July 7, 2015. https://hbr.org/2015/07/what-every-manager-should-know-about-machine-learning

[4] Steven Levy. “Inside Amazon’s Artificial Intelligence Flywheel.” Wired. February 1, 2018. https://www.wired.com/story/amazon-artificial-intelligence-flywheel/

[5] Jeffrey Dastin. “Amazon scraps secret AI recruiting tool that showed bias against women.” Reuters. October 9, 2018. https://www.reuters.com/article/us-amazon-com-jobs-automation-insight/amazon-scraps-secret-ai-recruiting-tool-that-showed-bias-against-women-idUSKCN1MK08G

[6] Kaya Payseno. “Avoid Amazon’s 3 Biggest AI Recruiting Mistakes.” Smartrecruiters.com. October 16, 2018. https://www.smartrecruiters.com/blog/avoid-amazons-3-biggest-recruiting-ai-mistakes/

[7] Susan Poser and Sharad Sinha. “How Machine Learning Can Improve Recruiting.” Oracle Profit Magazine. February 2018. https://blogs.oracle.com/profit/how-machine-learning-can-improve-recruiting

[8] Kumba Sennaar. “Machine Learning for Recruiting and Hiring – 6 Current Applications.” TechEmergence. December 15, 2017. https://www.techemergence.com/machine-learning-for-recruiting-and-hiring/

Previous:

Machine Learning Is Already Changing How We Buy Online, But Will It Transform How We Think Of Commerce?

Next:

Training Robots to Curate Individualized News

Student comments on Machine Learning for Recruiting at Amazon – Challenges and Opportunities

  1. I was shocked when I learned that the machine-developed algorithm was biased against women and penalized resumes that included the word “women” and favored candidates who described themselves using terms more commonly found on male engineers’ resumes such as “executed” and “captured.”

    I wonder if Amazon could keep the ML algorithm but manually remove that particular predictor variable, along with others that suggest similar types of biases.

    1. I agree with the comment above and think this is a good example of a situation where a machine-human combination would be more effective than a machine on its own. I think the issue of getting past prior biases in the human recruiting process is a challenging one, and I wonder if the value here is more in the machine drawing out interesting correlations rather than actually making the hiring decisions (or slimming the field) on its own. If the machine is able to provide visibility into what correlations are driving its specific recommendations then 1) the machine may bring to light important indicators of future performance that were previously unidentified by humans; and 2) a human could then make judgements on whether or not those correlations are valid criteria for making a hiring decision.

    2. I agree that this is a huge issue. It underscores the fact that a ML model is only as good as its training data! I think this issue can be solved by creating new metrics by which to grade people. As far as how we can evaluate machine bias, Amazon could create models that evaluate their models (kind of meta, I know), or use humans to evaluate the predictive models.

  2. You pose an interesting question: Is machine bias equivalent to human bias? This case would seem to indicate that the answer is “yes.” However, I wonder what data was used to identify successful candidates. While it is clear from the article that Amazon used information from applicants resumes, it is unclear how those inputs made women appear to be less capable candidates than men. Was Amazon using supervised machine learning and was the historical data that identified successful individuals biased? This is a case where bias from machine learning probably reflected bias within the firm or from society as whole. We should also be careful to ensure that machine learning allows us to correct our biases instead of perpetuating them.

Leave a comment