Machine learning collides with human resource management

Preface

In recent years, there have been many major breakthroughs in the field of machine learning, and there is a large and dynamic market for human resource management service products driven by artificial intelligence technology. More and more enterprises and government agencies are gradually thinking about applying machine learning technology to human resource management, making correct and effective decisions through neural networks, and accurately predicting the results of human resource management.

This paper introduces the application of machine learning to human resource management research from four aspects, including technical difficulties, introduction to human resource management decision-making system, system design methods and system security.

Technical difficulties

In 2019, CEOs of 20 major U.S. companies conducted a related workshop, and the results showed that the application of machine learning technology faces unique challenges in the field of human resource management. Developing a valuable HRM decision-making system is not only a technical challenge, but also requires barriers to measuring the inherent complexity of HRM outcomes, as well as data, ethical, legal constraints and possible negative impacts on affected employees that are difficult to address. Decision-making in human resource management needs to avoid selection procedures that are easy to challenge the law, or that employees or other stakeholders consider controversial.

The summary has the following aspects:

• how to establish and oversee a range of research and development projects that explore the application of machine learning in human resource management;
• how to effectively develop NLP-based decision support systems;
• how to test decision support systems to confirm that they are safe to use in decision-making;
• Once the system has been developed and tested, how to successfully convert the system to acceptable use.

Introduction to the Human Resource Management Decision-Making System

The implementation of the human resource management decision-making system faces the following challenges:

• Should the system automate decision-making, provide input to human decision-makers, or otherwise interact with the decision-making process?
• What kind of inputs do human decision-makers need, and how effective are candidate machine learning systems at delivering those inputs?
• What are the risks of different types of decision support, given the level of functionality currently available for different candidate systems?

The following diagram illustrates a framework for conceptual design and machine learning systems for human resource management. The idea behind the framework is that the design of the system is inseparable from the highest priority goals of the system. HRM goals help designers choose from the many possible ways in which machine learning can support the HRM decision-making process. The design of the implementation in turn influences how the system is evaluated. For example, systems that automate decision-making can be evaluated based on their accuracy or other important criteria, while systems that provide input to human decision-makers must make judgments based on the accuracy of the inputs and how they affect the overall decision-making outcome. If the system does not meet the safety standards, the implementation design must be modified until the designer is able to obtain a system that is valuable to the human resource management objectives and is able to meet the safety parameters.

System Design Methodology

In the early stages of development, there are many design options that can incorporate machine learning-based inputs into decision-making. Designs differ in time (e.g., before or after a human makes a decision) and degree of influence (e.g., recommending an option or directing attention to important functions). Here we will focus on the five main design implementations of machine learning decision-making systems:

1. Decision. Machine learning systems score human resource management records and automate decisions without human decision-maker involvement.

2. Referrals. Machine learning systems provide recommendations to human decision-makers as additional inputs.

3. Scoring. Machine learning systems provide scores to humans as additional input.

4. Summary. Machine learning systems automatically summarize for human decision-makers.

5. Audits. Machine learning systems flag anomalies for human decision-makers to review as part of the audit process.

The design process begins with the identification of prioritized goals for the machine learning system, with different combinations of goals requiring different design implementations, as shown in the table.

These objectives also point to potential measures of the effectiveness of the evaluation process. For example, if the goal is to reduce workload, the system should reduce the number of human decision-makers or the amount of time they spend recording scores, and if the goal is to improve human decision-making, the system should help improve the quality of decision-making, and measure decisions through evidence to better contribute to important HRM outcomes.

Machine learning systems that automatically summarize narrative records can serve as a modality for decision support. Most of a person's HR records are divided into two types of free-form text and people attributes. Free-form text, such as task lists, job descriptions, and summaries of key accomplishments. Personnel attributes are pre-quantified, interpretable, and management-useful data, such as years of experience, merit order, or promotion test scores. While the latter type of information is easier to process and use in models or visualizations, the former type of information is also needed to make well-informed HR management decisions.

Management's handling of decisions requires a thoughtful review of records and a manual review or scoring process performed by experienced personnel. Of the various design implementations considered to support manual review, "summary" is the most versatile. This is the only design that is moderately or highly aligned with all HRM goals. Automated summaries are useful for providing feedback, increasing transparency, and improving the accuracy of human decision-making, and they are at least somewhat useful for standardizing and reducing manual workload. At the same time, the summary implementation maintains a high degree of manual control over the decision-making process, so it is more likely than other designs to meet safety standards. In fact, a summary highlights the textual elements that the system considers important, and as such, it is an explanation of the system's decisions. Therefore, summarization can be a useful aid to help managers understand the model output in other design implementations.

System security

Human resource management decisions are a vital force that affects the future of an enterprise. Therefore, it is essential to apply the principle of "first and do no harm" when making significant changes to the decision-making process. As investment in machine learning increases, a plethora of research and policy documents aim to provide normative guidance for the responsible and ethical use of machine learning (and AI more broadly).

For example, the current rules and frameworks that protect the privacy of members will continue to apply to any development project. During development and deployment, there are three principles that are particularly relevant to test systems, which require machine learning systems to be accurate, fair, and explainable:

Accuracy means that the machine learning system or the model it contains correctly predicts the outcome of interest with a high probability.

Fairness means that machine learning systems treat subgroups equally.

Explainability means that humans can understand the factors and relationships that lead to the outcome of a machine learning system.

These safety standards sometimes conflict with each other. To increase fairness, designers may impose limits on the system, reducing its accuracy or explainability. To increase explainability, system designers may use modeling approaches that are more explainable (but less flexible), which can compromise accuracy and fairness. Testing must include a design that balances accuracy, fairness, and explainability to meet human resource management objectives and legal and ethical constraints.

With regard to fairness, it is important to note that there is no single definition of fairness, and it is often impossible to satisfy the fairness of the type of competition. As a result, the agency must choose a definition to move forward with the test. A distinction is made between procedural fairness, which ensures that the HRM process or algorithm treats members of different subgroups equally, and outcome fairness, which examines the model or process results for bias.

Finally, explainability is essential to achieving human resource management goals, as people may ignore or abuse systems if they don't understand how they can contribute to better decision-making. In addition, defining explainability is inseparable from the target audience, as different types of users require different levels of interpretation. Designers can consider using models that are inherently explainable to increase interpretability, and they can also conduct human-in-the-loop testing to assess how well people understand the functionality of the system.

brief summary

This paper briefly introduces the research on the application of machine learning in the field of human resource management from four aspects: technical difficulties, introduction to human resource management decision-making system, system design methods and system security. It is hoped that it will be helpful to readers who want to get a first look at this research.

For more information, please click Full-Scenario Live Streaming Solution - Aerospace Cloud Network Solution