Machine Learning Times
Machine Learning Times
Elon Musk Predicts Artificial General Intelligence In 2 Years. Here’s Why That’s Hype
 Originally published in Forbes, April 10, 2024 When OpenAI’s...
Survey: Machine Learning Projects Still Routinely Fail to Deploy
 Originally published in KDnuggets. Eric Siegel highlights the chronic...
Three Best Practices for Unilever’s Global Analytics Initiatives
    This article from Morgan Vawter, Global Vice...
Getting Machine Learning Projects from Idea to Execution
 Originally published in Harvard Business Review Machine learning might...

8 years ago
Model Your Workforce with Predictive Job Maps


Over the course of their careers, many employees develop skills and new interests, and move between roles in their organizations. Over time, their movements build pathways between job roles, creating a “map” of what kinds of people move between which roles. We see basic maps like this in workforce planning and diversity studies.

The map becomes predictive when we overlay employee-specific predictive models onto the map. These models transform the map from a static planning tool into a living representation of the organization and its potential.

Pasha Roberts, co-founder and Chief Scientist at Talent Analytics Corp., fielded questions from Data Informed about how organizations can apply predictive analytics to map these movements and determine which types of employees are most likely to change roles, which roles best suit individual employees, and how this knowledge can improve hiring and employee performance.

Data Informed: What kinds of data make up a predictive job map?

Pasha Roberts: Each node in the map is a role, for example, “Underwriter I” or “Inside Sales Representative.” Then, we map the traffic between each role: hires, promotions, demotions, transfers, terminations, and actual performance in the role. This gives us a set of history and probabilities for how people really move and perform through the organization, which is not always how HR has planned it. This emergent order usually brings a lot of insight, but it is nothing new. People have been doing this with Markov chains for a long time.

Pasha Roberts, Co-Founder and Chief Scientist, Talent Analytics Corp.

At each job node, we have a variety of data about the role: survival, cost, benefit, KPI performance, and employee lifetime value. We also store information about the employees that have been in that role, such as aptitude, skills, and tenure.

Most importantly, each role has predictive models attached, which predicts employee-specific KPI performance and survival. Examples of KPI performance include sales achieved, errors made, miles driven, merchandise sales, customer satisfaction scores, medication errors, any “real” business performance metric that can be measured. This “tree of models” is what makes it work.

What kinds of insights can a predictive job map deliver?

Roberts: Two kinds of decisions come from this map: macro and micro.

At the micro level, we can identify attractive paths for candidates or employees through an organization, years before the person has developed the skills or experience to take the later roles. An attractive path is defined as a role in which they have a high probability of being successful in the role: If a candidate applied for a customer support role but the models find he would be more likely to stay and to produce more value as an inside sales rep, we can put the candidate there instead. If we face a long-term need for regional sales managers, we can hire and promote sales reps that tend to become great managers.

While employees benefit from having roles in which they have a high probability of success presented to them, the main reason to do this is to save and make money for the organization. Everything about the network of models is tuned to maximize overall KPI business achievement.

At the macro level, we can look at workforce planning with much more information. Instead of just knowing that we will need more sales managers in 2018, we can create hiring/promotion models that fill those positions with better people.

What are some other ways in which an organization can put these insights into practice?

Roberts: At the employee level, a map of models will make smarter choices for the entire organization. For example:

  • Recruiting internal and external candidates with the highest chance of high KPI performance and low attrition.
  • Quantitatively identifying employees with “high potential” for management.
  • Identifying previously-unseen career moves for unproductive or unhappy workers.

At the workforce level, a map can be tuned to:

  • Perform traditional workforce planning with far more knowledge of employee lifetime value and capacity.
  • Design future staffing patterns years ahead by changing hiring models today.
  • Identify patterns to quantitatively “export corporate culture” to new sites and regions.
  • Run simulations and experiments to determine the impact of different scenarios.

I tend to agree with Holger Mueller that “insight” and graphics are interesting, but “real” predictive analytics should quietly, automatically be part of daily operational decisions. In our upcoming presentation, we have some pretty cool interactive graphics, but they are just the means to describe the problem. The predictive solutions to the problem are baked into much simpler decision tools for recruiters and managers.

Is a predictive job map more effective for gleaning predictive insights regarding a single employee, group of employees, or an entire workforce?

Roberts: The micro and macro components of the map work together. One informs the other. Individual employee movements teach the map about what tends to work. An exception, say an employee going from the call center to underwriting, can act like a mutation that identifies a whole new possibility for growth. With aptitude and skill information in the map, models can compute uncharted paths for growth by clustering similar roles. And we love the idea of predictive models charting new promotion and growth paths for employees.

But an organization rarely operates on the workforce as a whole, except in a reorganization or layoff. Usually the workforce is grown and directed one person at a time. The map lets us move people through the corporate lattice with a larger, evolving principle in mind, perhaps like a bonsai tree is formed over many years.

Is this tool more useful for determining an employee’s potential career path within an organization, or is it equally useful for identifying those at risk for leaving the organization, and perhaps likely destinations for employees at risk for leaving?

Roberts: It is useful for both. Survival modeling within a single role, a single node, on the map can identify who is likely to leave. With enough data, we can identify why they are likely to leave. You don’t need a full map just to know that.

The map comes into play to deal with someone who is at risk of leaving – usually we want to save our investment in that person. In that case, we can compare that person’s aptitude and skills to hundreds of other job models and find out where she could do better. By better, I mean last longer, be happier, and produce more value for the organization, as well as more job satisfaction for the employee.

By far, the most profitable way to use a predictive tree is in recruiting, before you hire someone. You save money by avoiding bad hires, the quick burn-outs or flights to competitors. And you create value by selecting people with higher employee lifetime value (eLTV). The candidate base is big, diverse, fluid and open, and far easier to allocate than existing employees. It costs a lot to get just one employee on board and up to speed. You want to spend that money wisely.

We work similarly to banks that predict the probability of default before extending credit. It makes no sense to extend credit and then predict probability of flight risk after the creditor is on board.

If an employee or group of employees is determined to be at risk of leaving, can the map help predict effective interventions, such as offering promotions, raises, training, etc.?

Roberts: Within a general flight risk number, we strive to understand types of termination: Is it because the employee wants bigger things, more money, or because they can’t do the job? These can lead to different interventions. Certain terminations are not regrettable, and we just can’t find a next-best role for the person. Again, this points to hiring predictively in the first place.

We see that different kinds of people respond differently to training, pay raises, or increases in status, for example. Uplift modeling shows us that people value these incentives differently, so our interventions can match. For example, some people are well suited to work at home, others need the discipline and interaction of an office. Some people value the political value of a bigger job title, and are also well suited to perform in that title. And some people value learning with courses, time to self-teach, or go to conferences.

In addition to identifying likely career paths and potential flight risks, can a predictive job map help with other aspects of HR? Can it inform decisions around benefits, for example?

Roberts: Just knowing a potential path for someone doesn’t do much for anyone. Likewise, just knowing that an employee has a 42 percent chance of departing within a year doesn’t do much on its own. It has to fit into a bigger picture.

Benefits is an interesting case. Guidelines for compensation are well-laid tracks that will take years to move out of strict general formulas. Many comp plans still follow very simple ladder-based structures that no longer represent how modern organizations behave. Instead, predictions and metrics can work around that aspect with discretionary bonuses and other perks that actually appeal to a specific employee.

A large amount of time and myriad unforeseeable and potentially impactful forces and possibilities exist over the course of an employee’s career. How do these affect the accuracy of predictions based on a job map?

CONTINUE READING: Access the complete article in Data Informed, where it was originally published.

Scott Etkin is the editor of Data Informed. Email him at Scott [dot] Etkin [at] wispubs [dot] com. Follow him on Twitter: @Scott_WIS.

Leave a Reply