9 February 2019

When governments turn to AI: Algorithms, trade-offs, and trust


Artificial intelligence can help government agencies solve complex public-sector problems. For those that are new at it, here are five factors that can affect the benefits and risks.

As artificial intelligence (AI) and machine learning gain momentum, an increasing number of government agencies are considering or starting to use them to improve decision making. Some examples of compelling applications include those that identify tax-evasion patterns, sort through infrastructure data to target bridge inspections, or sift through health and social-service data to prioritize cases for child welfare and support. They enable governments to perform more efficiently, both improving outcomes and keeping costs down.

The most pressing aspects of adopting such solutions are generally well known. Algorithms should be accurate and consciously checked for unintended bias.1 Others are less so. Algorithms must be stable, meaning that small changes to their input don’t meaningfully change their output. They should be explainable, especially in the public sector, where myriad stakeholders will review every step. And to ensure successful adoption, public-sector users should pay particular attention to how AI solutions are deployed, given public-sector managers generally have less authority and operational control to compel adoption than private-sector ones. While all these factors are relevant to every public-sector entity, they aren’t necessarily relevant in the same way.

Getting the right balance is essential not only to minimize the risks but also to build a proper business case for the investment, and to ensure that taxpayer dollars are well spent. Below, we’ll explore each of these five dimensions—accuracy, fairness, explainability, stability, and adoption—as they apply to the public sector.

Accuracy

When it comes to algorithms, public-sector users could measure performance in terms of better decision making. Since there are typically many possible measures and probabilistic outcomes, it’s unlkely that an algorithm will forecast every one of them precisely. Users could start with identifying which ones are most likely to lead to the best decisions for the situation. We recommend focusing on two or three measures that truly matter for the specific use case. Consider the following examples:

Deciding which individuals receive rehabilitation treatment.Correctional officers or social workers at prisons may prefer the algorithms to reduce the number of false negatives—high-risk individuals falsely classified as low risk—relative to false positives—low-risk individuals falsely classified as high risk. That’s because the potential impact of missing a high-risk individual could be higher likelihood of recidivism while that of misclassifying a low-risk individual would be additional programming.

Deciding where to focus tax audits. Tax officials may want to optimize for focusing on only the most likely tax evaders—given the potential consequences of falsely tagging someone as a high risk for evasion.

Deciding which students get scholarship money based on probability to graduate. When the rank order of students determines scaled scholarship amounts, the order in which students rank could matter more than the absolute probabilistic score that the individual student receives from the model—in this instance, the likelihood of graduation. In such cases, school administrators would care most about predicting the correct ranking order of the students than the accuracy of the probabilistic outcome by itself.

One word of caution: ensure that a clear baseline accuracy for decision making exists before implementing an algorithm, whether based on historical human decisions, rudimentary scoring, or criteria-based approaches that were being used. Knowing when the algorithm performs well and when it does not, relative to the baseline, is helpful both for making a case to use it as well as to establish incentives for continued improvement of the algorithm.

In our experience, machine learning can significantly improve accuracy relative to most traditional decision-making processes or systems. Its value can come from better resource-allocation decisions, such as matching the right types of rehabilitation programs in a corrections facility to the prisoners most likely to benefit from them. But it can also be valuable for improving efficiency, such as helping public-health case workers prioritize the right cases, as well as effectiveness, such as knowing which school programs are most effective at minimizing drop-outs.

Fairness

There are many ways to define a fair algorithm, or “algorithmic fairness.”2The notion reflects an interest in bias-free decision making or, when protected classes of individuals are involved, in avoiding disparate impact to legally protected classes.3 There is extensive literature on bias in algorithms and how this could manifest. Common issues include some kinds of bias in the data sets and distortions in the algorithm’s analytical technique—or in how humans interpret the data.

A critical first step is to establish what fairness means in the specific context of the use case—that is, what are the protected classes and what are the metrics for fairness. There are a few ways to measure and address fairness, not all of which may be equally effective in each instance:

Willful blindness. One approach that is commonly used is to build a kind of blindness into the algorithm, so that it treats subgroups the same regardless of traditional distinctions between them, such as race, gender, or other socioeconomic factors.

For example, if a school uses an algorithm to identify students at risk of dropping out, educators could deploy a model that uses gender-masked or gender-neutral records to identify those at the greatest risk. Yet even that kind of approach can be naive if it doesn’t account for cross-correlated variables—such as postal codes that could imply race, education level, or gender. Such an approach could lead to unfair outcomes or cause issues with the sample data used to train the model itself. It ends up creating an algorithm that is merely unaware without any consideration to fairness.

Demographic or statistical parity. Another way to address fairness is to ensure statistical parity in the decisions being enabled or in the outcomes—for example, by selecting an equal share of people from both protected and nonprotected groups. One way to achieve this would be to set different thresholds for different groups to ensure parity in the outcomes for each group.

An example of the latter would be an algorithm written to apply different credit-score thresholds for different demographic groups, in order to select the same proportion of applicants from each. However, this approach requires someone to constantly verify and modify the thresholds—and often may not account for underlying differences in the subgroups. It is usually effective only when someone cares about a single measure of fairness, in this case, an equal share of loan-approval outcomes across gender types.

Predictive equality. Possibly the most balanced approach to address fairness is to not force it in the decision outcome, but rather in the algorithm’s performance (or accuracy) across different groups. In this definition, fairness means that the algorithm is not disproportionately better or worse off in how decisions are being made for specific subgroups. That means, for example, that the error rates or prevalence of false positives or false negatives for each group is the same—while accounting for variations in the underlying population. In our loan-applicant example, this means that we may not approve an equal share of loan applicants across genders, but the percent of approved applicants who end up defaulting (that is, the false positives) would be the same across genders. In other words, we are not disproportionately favoring or affecting either gender as we are making the same rate of mistakes or errors in our selection.

Steps to debias algorithms include the following:

Identify the specific subgroups or protected classes that are relevant.

Identify the set of metrics that define fairness, and any implicit hierarchy within those, if you have more than one.

Evaluate the training data set for adequacy across subpopulations or protected classes—and collect more data where needed.

Identify features such as zip codes, income levels, or other socioeconomic data that are correlated with the protected-class variables or groups—and either remove or adapt them. Advanced methods could use machine learning to identify how biased the model is; as an example, if removing race from a model does not change the outcomes at all, then potentially other variables are strongly cross-correlated.

Evaluate fairness outcomes for different model types, across different time periods. Consider if specific models for different classes or subgroups may be needed (or thresholds or adjustments may be required).

We should note that fairness may come at a cost of lower accuracy. For example, we may find an algorithmic model is highly accurate for a population overall—but not for some subsets of the population where there is less data. In the case of correctional systems, changes in demographics of the prison population could render models of behavior moot, if they are based on historical data. Additionally, they could have more accurate outputs for historically dominant groups and potentially a higher or unfair error rate for others.

There can be a trade-off between higher overall accuracy at the cost of poorer, less-fair performance for some and more fairness (by removing certain features) at the cost of reducing overall accuracy. For example, if certain variables in the underlying data, such as postal codes, are correlated with race in certain geographies, then adding postal codes to the data set used by a model to be more accurate could inadvertently introduce racial bias. Hence, in picking the right model, it is important to look at how algorithms score across the five dimensions we have outlined here.

Explainability
Sidebar
Privacy, integrity, and vulnerability

Easily explained algorithms can be critical in encouraging the adoption of an AI application, ensuring that stakeholders understand how and why decisions are reached. In our experience, AI and machine learning are most valuable when used to support, and not substitute for, human decision making—and to enable the same humans to understand the rationale behind the algorithm’s recommendations. In our experience, just making a real person available to engage with those affected by consequential decisions can make a difference, even if the decision isn’t changed. Many public-sector systems are already designed to enable this, such as judicial hearings and public-comment periods around policy decisions. This combination of “human plus machine” can actually often make substantively better decisions than the machine or the human on their own (see sidebar, “Privacy, integrity, and vulnerability”).

This is particularly relevant regarding decisions to allocate a scarce resource, such as when an algorithm’s output helps select a limited number of applicants for scholarships, grants, or permits. In extreme cases, a black-box AI application—one that isn’t or can’t be explained—can potentially cause more harm than help. Machines can make errors and reach rigid conclusions, especially in narrow borderline situations. For example, an algorithm might deny a loan for an applicant with a credit score of 728 when the cutoff is 730. People can only correct errors or make exceptions when they understand how the machine makes decisions.

Like fairness, explainability can also lead to difficult trade-offs. Simpler algorithms using rules-based heuristics or decision trees may be easier to explain, but more nuanced and complex algorithms might be more accurate or less biased. The determinative question is whether it’s more important that people understand the rationale behind a decision or more important to be accurate.

The answer is contextual. In some countries, for example, various credit-scoring systems4 can have wide-ranging implications for an individual’s ability to get a loan. In such cases, a more explainable algorithm would give applicants an opportunity to improve their input variables, such as avoiding late payments, to influence their final scores over time. In contrast, if an algorithm accurately identifies patients with high risk of cancer, patients are unlikely to care if the algorithm is easily explained.

Organizations can also consider moving to more complex algorithms once the user base becomes more familiar with and trust is built in the more explainable models.

Stability

Over time, the performance of most algorithms grows unstable, primarily because they were developed using data collected in a world before algorithms were used to make decisions. For example, changes in court-sentencing patterns could increasingly change the type and characteristics of offenders entering a prison. Those changes in the offender population eventually render any predictive algorithms increasingly irrelevant for the current population. Traditional risk-scoring systems or even human decisions face the same obstacles.

This is particularly important in the public sector, where many external factors come to play in decision making. To continue our criminal-justice-system example, laws get relegislated. Judges change how they sentence and for what crimes. And prison populations change as age, economics, social environment, and other demographics evolve. Just introducing a machine-learning model to classify offenders can lead to behavioral changes that affect the underlying data set. For example, if correctional staff prioritize rehabilitative treatment toward high-risk offenders, then the mere fact that an offender is flagged as high risk can actually lower his or her risk over time. That’s a trend that a data set created before the emergence of machine learning would not have seen. Such changes can create shocks to the system that render historical data less capable of predicting the future, invalidating traditional heuristics or decision-making rules.

To estimate the frequency at which models should be refreshed, users must understand the speed at which algorithmic performance degrades. One way to do this is to test its performance using backward-looking data over different time spans. If the model performs great on test data that lapsed a year ago but not on data that lapsed two years ago, then retraining the model somewhere between a year and two years will likely help avoid degradation. Ideally, organizations would use such information to develop a cadence of regular testing and retraining to continuously update and rebuild their heuristics. However, models may also need to be refreshed in the wake of any major changes to an underlying data set. These might be internal changes, such as the implementation of new policies, or external, such as new legislation.

Adoption

When we think of the potential impact of AI, we think of three big pieces of work: developing the model or algorithm, deriving insight from its output, and adopting its output or recommendations. In the end, a great machine-learning model, by itself, is not enough. It often needs to be wrapped in an intuitive user-centric experience and embedded into work flows, with the use of design thinking and with frontline employees to spur adoption.

Machine-learning algorithms are prone to rejection for the same reasons they deliver great results. That is, they can generate accurate but counterintuitive insights due to the large number of variables and data they use. They go against the grain of traditional heuristics. They challenge the ways things have traditionally been done. And they often require people to give up familiar tools and methods.

Thus, it is critical to plan for, and to incorporate approaches to encourage, adoption from day one. These might include bringing target users into a model’s development process from the beginning—or at least soliciting frequent reviews and input along the way. It might also include designing a straightforward way to deliver and consume the model’s insights. At the corrections agency mentioned above, for example, officials experienced an excellent response to adoption because corrections officers were so excited about the insights the applications offered—and because of the intuitive user interface, which consolidated disparate sources of data (including paper) into one easy-to-use front end. That made their work less tedious, thereby encouraging them to use both the analytics and the tool.

While important, adoption is where typical analytics teams struggle, whether internally in public-sector agencies or in external partnerships with vendors. Proper adoption requires end-to-end expertise, from use-case articulation to model development, tool development (insight delivery), and, ultimately, change management and operational rollout. The need for these cross-functional skills and expertise makes this last mile often the most challenging one.

No comments: