Machine learning is everywhere. Whether you realize it or not, your life is already impacted by the algorithms and recommendations created by machine learning applications. Companies like Stitch Fix, Uber, Instacart, IBM Watson, and Tesla all use machine learning to understand their users’ behavior and to deliver better service. Besides these consumer companies, HR service and product providers such as Wade & Wendy, RiseSmart, and Randstad are also using machine learning and artificial intelligence to help narrow down job candidates or to match appropriate jobs to qualified candidates looking for those jobs. I discussed Putting Machine Learning to Work in HR in depth during a recent #SmartTalkHR webinar. You can watch the webinar in its entirety here.
The Stitch Fix use case
For every company using machine learning, there is a process that begins with input of the machine learning algorithm and a calculation made by the algorithm that leads to a specific output. For Stitch Fix, they collect the style preferences, sizes, and buying habits of their subscribers to continue to adjust what clothes are sent to each person, based on stated preferences as well as actual choices and buying habits. That information is then combined with available inventory and collaborative filtering that allows Stitch Fix to make choices for one person based on the decisions of another, similar customer.
What’s most interesting about Stitch Fix, and why I chose it as a machine learning application in HR, is that the company doesn’t just rely on the output of the computer to make decisions for their customers. The list compiled by the computer is then handed to a personal stylist who can use the machine learning algorithm as a tool. They’ve done a really good job of letting a machine do what a machine does best — digesting a large volume of data and picking out the patterns that would be difficult for a human to determine. Then, they let the humans do what people do best – building relationships, giving context, making the personal connections with the customer, and then using instinct to make the final decisions.
The RiseSmart Insight™ use case
So, how does this kind of model translate for HR? RiseSmart is using this same model of tech and touch to help match job seekers to open jobs based on experience, skills, and preferences. Using machine learning, SmartMatch™ combines input from the job seeker including preferences, resume, and other specific information which is matched against job listings from across the internet. Using patented ontology technology, SmartMatch understands the intent behind the job search. For instance, if someone lists Visual Designer as a preferred job title, the technology determines what that titles includes, and what it does not include. The ontology actually allows the application to conduct a smarter search that goes beneath the surface and learns about the job seeker’s preferences and best matches based on which postings receive clicks, which ones are bounced from after a short time, and which ones are ignored. Then, following the model of tech and touch, a personal job concierge reviews job listing results and hand-picks those jobs most suited to the individual candidate, based on their personal knowledge of the person.
Using survival analysis in recruiting
The use of machine learning is becoming increasingly prevalent as more businesses are using these algorithms to make their businesses work better and smarter. The use cases for machine learning are growing as are the types of metrics available from these algorithms. At Randstad, we use machine learning to improve our recruiting processes and to redistribute recruiters and resources, when appropriate. One type of machine learning process we use is survival analysis.
Survival analysis is a technique for predicting time to event based on historical data. In medicine, the technique is used to establish prognosis for serious diseases, based on risk factors. By looking at the time to fill problem in recruiting in the right way, we can take advantage of the same mathematical approach. Consider a job as a patient and the time to event as the time to fill the job instead of the time from diagnosis to death. For risk factors, we use various features of the job along with market data. We want to understand the probability that a job will remain open beyond its target time to fill date. In this case, our patients are jobs, and we want them to die quickly. We want the survival time to be short. Although we are thinking about survival in terms of days, hopefully, instead of years, we can take advantage of the survival analysis technique developed for medical applications.
The data used in this analysis is historical data from our customers on job features and accompanying time to fill along with job market statistics from Career Builder capturing what the job market looks like for a position in this industry. The job market data includes:
• How many open jobs in a particular field
• How many candidates looking for work in that area
From this data, we built a survivor analysis predictor. Then, as new jobs opened, we could assess in real-time the probability that the job would remain open past its target time to fill date. Based on this analysis, we make business decisions and apply additional resources to jobs that are at high risk. Some of those are obvious, but some are counterintuitive.
Using the output of this predictor, we are able to assign a risk category to each of the open jobs – high, medium, or low risk that the job will miss target time to fill. By matching open jobs by risk to the recruiter loads, we can see who has a large portion of job requests in the high or medium risk segments and shift some of those job requests around or assign additional resources there. We can also sort by geography and apply the same remedies.
Talent advisory toolkit
In addition to applying survival analysis to help improve our time to close job requests, Randstad is currently developing a Talent Advisory Toolkit to help our clients gain visibility into the data they need to fill a job, including:
• Best sources for the job
• Time to fill information
• Market salary data
• Availability of contract workers
• Geographic data
The difficulty in getting accurate and complete data rests on our ability to aggregate job titles. This is a similar challenge as the one faced by the developers of SmartMatch. There are a lot of different job titles that represent the same thing and the challenge is to gather all the job titles that are similar. This is a natural language processing problem and requires a semantic search that incorporates both the meaning of the job titles as well as the actual words used.
Applying advanced mathematical techniques to solve real-world problems and improve our ability to streamline processes and improve our products and services is an area that will continue to grow and evolve. As HR technology companies continue to offer solutions unique to human management challenges, AI and machine learning algorithms will become part of the tools and processes of HR departments in businesses of every size.
If you’re interested in a deeper dive into machine learning, its applications in the world around you, how it works, its applications in recruiting, and the future of machine learning, I discuss those topics in detail in my #SmartTalkHR webinar, Putting Machine Learning to Work for HR available for viewing here.
Summer Husband is Senior Director, Data Science for Randstad Sourceright. Through deep analysis and visualization, applying predictive analytics and machine learning, Summer is able to support clients’ preemptive decision making and deliver continuous and measurable improvement across all aspects of talent acquisition