Machine Learning and AI in human relations departments

towards-data-science

This post was originally published by Mitchell Telatnik at Towards Data Science

A New Age in Productivity or Removing the ‘Human’ from HR?

Deep learning and AI have been drastically changing industries such as healthcare, financial services, and retail with many companies welcoming new technologies. However, Human Resources (HR) departments have been met with more challenges in integrating intelligent systems into their workflows.

HR Departments are tasked with managing the organization’s employees — hiring, firing, resolving disputes, payroll, benefits, and more. Many of these tasks seem ripe for automation with machine learning, however, they are also often subjective, and handing over the reigns poses interesting ethical challenges.

The hiring process is laborious and expensive. From reviewing resumes, interviewing, and training new employees, hiring new employees can carry a large cost to organizations outside of the new employee’s salary. This cost is generally worth it, however, because making the wrong decision can cost even more money if the employee must be let go and the process started again. Not only do the costs of hiring need to be incurred a second time, but lost production and the time it takes the new employee to ramp up to full production will also be present.

Because of this, many companies have looked towards deep learning to provide solutions in reducing the cost of hiring employees and increasing the quality of employees hired. However, these attempts have not always gone as planned.

From 2014 to 2018, a team at Amazon built systems to review applicants’ resumes in an effort to streamline the process of recruiting top talent. In order to train their algorithm, the team compiled a training dataset using resumes submitted to the organization over the prior ten years.

Amazon had hoped that this system would drastically reduce the time it took to identify top talent out of the applicant pool by automatically identifying the top x number of applicants. However, they later discovered that the system was favoring male over female applicants. This was because more male than female job seekers submitted resumes to Amazon, creating a biased and skewed dataset.

Creating unbiased hiring systems can be a difficult task. Since most companies rarely have exactly 50% male employees and 50% female employees, the model can often identify factors that it thinks are most telling of a good hire but are actually not considered by the hiring managers.

In order to create accurate hiring decisions and candidate ranking systems, care must be taken in assembling the datasets for training to eliminate unwanted behavior. In addition, it may be viable to hardcode the model to disregard certain features such as name, gender, and race.

HR Teams managing hourly employees have a daunting task when creating schedules. When your employees are not all full-time with a consistent schedule, scheduling conflicts often arise. Because of the unpredictable work schedule, a key function of HR managers (and often general managers) is managing time-off and shift-change requests.

If you ask many restaurant or retail store managers, scheduling and it’s related tasks often take up a large portion of their workday. However, deep learning systems are starting to take this burden.

Automated systems can analyze these requests and automatically approve or deny them based on predefined business rules on a personal level. For example, many organizations that employ part-time shift workers do not allow their employees to work 40 or more hours in a given week. If a shift-change request puts one employee above 40 hours for the week, the system would decline the request without any human intervention.

These systems become even more powerful when combined with predicted demand information. Accurately predicting when additional employees are needed and adjusting schedules and time-off requests accordingly can provide increases in the efficiency of employee management by saving on labor costs when demand is low and assuring adequate employees are working when demand is high.

While these systems can drastically improve employee management and reduce the workload for managers, it could hurt employee morale. Often times time-off and shift-change requests can be personal in nature. If an automated system denies a time-off request for an important event, the employee may grow resentment towards the organization.

For automated scheduling systems to work, it is important to thoroughly define business rules and ensure there is a way for employees to receive an override of the model’s decision through their manager.

Analytics teams have allowed organizations to harness their data in ways never before allowing companies to make more informed decisions than ever before.

When we think about business data, we often picture them collecting data on their customers to better understand them. However, many organizations are also collecting data on their employees.

Tracking key metrics on an organization’s employees allows HR departments to better understand their workforce. Tracking employees’ sentiment, productivity, and connection to the organization empower HR departments to better allocate resources and improve the efficiency of their workforce. In addition, it allows predictive analytics models to identify employees at risk of leaving the organization or likely to be promoted.

While workforce analytics may seem like an obvious choice for HR departments, employees may not have the same perspective. Distilling an employee into a series of numerical metrics can dehumanize the management process and make employees feel like nothing more than a number.

Machine learning and AI have promising applications in human resources. However, switching from human to machine-driven management can cause major issues in employee morale within an organization. Do these technologies create a new age in productivity or remove the ‘human’ from HR? I’ll leave that up to you.

Spread the word

This post was originally published by Mitchell Telatnik at Towards Data Science

Related posts