Artificial Intelligence is being deployed to address many human problems, most recently Google’s Duplex can make reservations on your behalf by talking to a human.
We have had some really interesting clients in the Human Resources (HR) space, a field dominated by human interaction. The main question we see from clients is how to use A.I. to do headhunting or job matching of candidates to roles. Today I want to walk you through the solution architecture for one of our clients in the HR space, and give you a sense for how A.I. can be deployed to automate and improve HR processes.
DREAM.ac builds teams with A.I., enabling more projects and startups to succeed. The motivating problem is simple: 68% of projects fail and 90% of startups fail. These failures can often be traced back to having the wrong team, weak leadership, or failing to raise funding. DREAM was built by founders who have already succeeded in the freelancer segment of the HR space, and so the requirements were driven by real-world experiences with the problems in the Startup industry, rather than a drawing on a napkin.
The problem the DREAM team (pun intended) wanted to address first was building successful teams. Back in 2000s, I worked on matching DVDs to people’s list of desired DVDs (see zip.ca). The technical name for that DVD assignment problem is a transshipment problem. The idea is to maximize everyone’s happiness, given the physical limit that a given DVD can only be mailed to one person at a time. Similarly, for DREAM we can only fill a job with one person, and a person can only do so many jobs.
To fit a person into a role, we use regression to give every possible person in our list of candidates a fitness score that represents the likelihood they will do well in the job, and by extension that the team will succeed in their goals. The person-to-job matching score is based upon the person’s past history of doing such jobs (experience) and the history of that type of person doing well in that type of role. Types of people and roles are "discovered" using clustering. The candidates are ranked by top-n so that the top fits for the job can then be considered for each position. The goal of this matching system is to maximize team success, rather than maximizing jobs filled. To cover as many general rules and exceptions as possible we architected a solution that uses wide and deep learning.
The solution ends up being a recommender system, which is near and dear to my heart, as that was the topic of my PhD thesis.
Some open source datasets can be quite useful in demonstrating the effectiveness of this idea in practice. Kaggle lists the following datasets, among others, to start playing with:
- Job Recommendation Challenge from CareerBuilder (2012)
- 19K Job Posts from CareerCenter (2017)
- 22K Job Posts from Dice.com (2017)
- 22K Job Posts from Monster.com (2017)
- 22K Job Posts from Naukri.com (2017)
- 50K Job Posts from Reed UK (2018)
As I have mentioned in a previous article, having a proprietary dataset is key to having a competitive advantage. Rather than building on these public datasets, DREAM is designed around private member data that is owned by platform users and shared with the company to provide training examples. The company only owns the derivative trained system, rather than the original data which users may choose to take back by closing their account.
Let’s look at the big picture of how teams get built, with the goal of having that team succeed.
The problem of building successful projects is a hierarchical one that requires the neural network model to understand (model) both teams and roles. We are continuing to build some really cool tech to address this gap in the way projects are assembled today by traditional HR approaches.
In conclusion, A.I. can be used to identify top candidates for a slot in a team according to their "fit" (score) for the role. This can be accomplished in the context of the larger process of assembling a team, and even in the broader context of developing a successful project.
To read more about DREAM.ac and the whitepaper, just click those links.
If you liked this article on building teams using A.I., press the follow button, and have a look at some of my most read past articles, like "How to Hire an AI Consultant." In addition to business-related articles, I also have prepared articles on other issues faced by companies looking to adopt deep machine learning, like "Machine learning without cloud or APIs."
Happy Coding!
-Daniel [email protected] ← Say hi. Lemay.ai 1(855)LEMAY-AI
Other articles you may enjoy:
