Predictive Analytics

Predictive analytics is the use of algorithms and machine learning techniques to forecast future events in real-time. Leveraging vast data sets, these algorithms are able to predict potential risks and costs, or even an individual’s future behaviour . Due to rapid advancements in technology and the emergence of big data, the use of predictive analytics has grown in prominence within various industries. Healthcare treatment recommendations are generated using predictive analytics, as well as assessments of candidates for hire, and it is even used by law enforcement to help anticipate potential crimes and criminals . As technology evolves, data analytics and artificial intelligence will continue to grow in capability and in potential applications. However, with predictive analytics becoming increasingly prevalent in decision-making processes that have direct and potentially life-changing impacts on people’s lives, ethical concerns regarding algorithmic bias, transparency, and data privacy are revealed .

Types

There are three types (or models) of predictive analytics. There are predictive models, descriptive models, and decision models . Each model has a specific function related to predicting and forecasting information. The most common type is a predictive model, but other forms of predictive analytics are used in different settings and contexts.

Predictive Models

Predictive models aim to make a prediction by comparing some object in one group to another object in a different group . By comparing the attributes and features of a specific object, one can make predictions about how another object will behave under a similar set of circumstances. These attributes and features form the variables of the model. Predictive models are built by feeding the model large amounts of training data, which is then used as a reference to generate predictions on other objects outside the training data.

Descriptive Models

Descriptive models differ from predictive models in that the focus is on groups rather than on the individual. Predictive models seek to make a prediction about the behavior of one individual, whereas descriptive models seek to categorize individual things into groups . Descriptive models generalize relationships between a large group of objects in order to categorize and organize objects by a variable. Descriptive models are typically used to make inferences about the behavior and preferences of groups of people.

Decision Models

Decision models are complex, and are generally used for predicting the results of a decision . Decision models are fed data from multiple sources, including other predictive and descriptive models, and use that to predict outcomes. Predictive models can predict individual outcomes, such as an individual's credit risk, but they cannot predict what would happen to the individual if they were denied or approved of a loan. However, decision models can make these complex predictions.

Uses

Predictive analytics have multiple applications and are used in a variety of contexts.

Healthcare

Graphic demonstrating how patient data gets fed into a model to generate predictive insights on patients.

Healthcare analytics refers to systematic use of health data and related business insights developed through applying analytics to drive fact-based decision making for planning, management, measurement, and learning in healthcare . Predictive analytics involves using similar methods to generate data predictions and other techniques for assessing predictive power . In the context of healthcare, it can be used to identify high-risk patients and provide treatment, reducing unnecessary hospitalizations or readmissions. Researchers Harris et. al. developed an analytical model to predict future patient behavior based on past behavior . This model provides an accurate prediction of no-show patients and assists clinics in developing operational mitigation strategies such as overbooking appointment slots and managing patients predicted as “no-shows” . Models such as these can be used for clinical planning and scheduling decisions to improve patient service at hospitals with optimal and unique solutions .

In healthcare, predictive analytics have been used to study Parkinson's disease . Additionally, some support its use for creating models that predict people who could have a greater chance of developing chronic diseases, thus helping them identify these illnesses earlier on, saving them time and money .

Human Resources

In the human resources field, predictive analytics and modeling can be used for things like forecasting openings within companies and anticipating which employees are a liability . The use of predictive analytics in human resources has recently been increasing in popularity . Companies use data and analytics to design, evaluate, and implement new management policies; this also means that the traditional methods of using experience, intuition, and guesswork to guide human resources strategy is falling to the wayside . Applying HR analytics in a firm can be a one-time effort that gradually becomes a newly overhauled approach to organizational management and it is not uncommon that these inspire more broad-reaching change to a corporation . Analytics must be rooted in understanding the data to be used and the context under which the data were collected to identify any potential biases .

A study conducted in 2019 of 4,800 individuals across different companies in different industries and determined that roughly one-quarter to one-third of all companies use predictive analytics in human resources . From this study, the industry that uses predictive analytics the most in human resources is financial services, with 32% of companies applying analytics . The Technology (software), Oil and Energy, and Healthcare and Pharmaceuticals industries all had over 25% of companies applying analytics to human resources .

Law Enforcement

In law enforcement, predictive modeling techniques have been used and referred to as PredPol (derived from the term Predictive Policing), as in the Santa Cruz California Police Department. Officers at this department state that it is used as a supplementary tool rather than a replacement for their normal rotations. Additionally, this PredPol system predicts solely based on crimes reported and not demographic or identifying information of individuals involved in the crimes in an attempt to reduce demographically based biases .

Ethical Challenges

Bias and Discrimination

Ever since the rise of the computer gaming industry brought back the resurgence of neural networks, experts have argued that deep learning is a highly effective way to train an artificial intelligence system . Designed to mimic the way a human brain thinks and makes decisions, a network of thousands or millions of individual processing nodes are connected together in a neural net, which enables an algorithm to train itself to perform a task given a prepared training data set . However, according to Barocas and Selbst from Cornell University and UCLA respectively, “an algorithm is only as good as the data it works with ." Zarsky, a professor and vice dean at Haifa University, argues that algorithms trained on biased data sets will not only inherit pre-existing biases from the aforementioned data set but also generate novel patterns of unfair bias and discrimination and reinforce these patterns in their decision-making processes . Some surmise an algorithm may interpret inequalities in historical data as sensible patterns, which support existing societal biases . Detecting and addressing unfair bias and discrimination in algorithms for predictive analytics is particularly difficult as it occurs due to unintended consequences from using the algorithm, not the purposeful actions of an ill-intentioned programmer..

Transparency

Some assert that transparency as an ethical issue is in opposition to other ethical interests such as privacy . Transparency in algorithms means that the algorithm should not only have its details accessible but also comprehensible to humans analyzing them; accessible information that is not decipherable is no longer useful . Many modern sophisticated artificial intelligence systems are trained via deep learning that is comprised of extensive neural networks reaching up to fifty layers . As each layer adds complexity, Sloan and Warner assert that the human comprehensibility of these networks is affected and thus the transparency .

Predictive Privacy

The term “predictive privacy” refers to the ethical challenges posed by the ability of algorithms to predict sensitive information about an individual using information derived from data sets of other individuals . In 2019, the Electronic Privacy Information Center (EPIC) raised this ethical concern in their official complaint to the Federal Trade Commission (FTC) against HireVue, a recruiting-technology company, stating that “the company’s use of unproven artificial-intelligence systems that scan people’s faces and voices [constitutes] a wide-scale threat to American workers ." Mühlhoff’s definition of a violation of predictive privacy is “if sensitive information about [a] person or group is predicted against their will or without their knowledge on the basis of data of many other individuals, provided that these predictions lead to decisions that affect anyone’s...freedom ." Some note that predictive privacy can be violated regardless of the prediction’s accuracy, especially when systems for data collection and processing are designed such that subjects cannot provide meaningful or informed consent .