Unlocking the Black Box – 3 Explainable AI Methods to Prepare for the AI Act

Data Science
Human-centered AI
Statistics & Methods

17. May 2023

Team statworx

The Hidden Risks of Black-Box Algorithms

Reading and evaluating countless resumes in the shortest possible time and making recommendations for suitable candidates – this is now possible with artificial intelligence in applicant management. This is because advanced AI technologies can efficiently analyze even large volumes of complex data. In HR management, this not only saves valuable time in the pre-selection process but also enables applicants to be contacted more quickly. Artificial intelligence also has the potential to make application processes fairer and more equitable.

However, real-world experience has shown that artificial intelligence is not always “fair”. A few years ago, for example, an Amazon recruiting algorithm stirred up controversy for discriminating against women when selecting candidates. Additionally, facial recognition algorithms have repeatedly led to incidents of discrimination against People of Color.

One reason for this is that complex AI algorithms independently calculate predictions and results based on the data fed into them. How exactly they arrive at a particular result is not initially comprehensible. This is why they are also known as black-box algorithms. In Amazon’s case, the AI determined suitable applicant profiles based on the current workforce, which was predominantly male, and thus made biased decisions. In a similar way, algorithms can reproduce stereotypes and reinforce discrimination.

Principles for Trustworthy AI

The Amazon incident shows that transparency is highly relevant in the development of AI solutions to ensure that they function ethically. This is why transparency is also one of the seven statworx Principles for trustworthy AI. The employees at statworx have collectively defined the following AI principles: Human-centered, transparent, ecological, respectful, fair, collaborative, and inclusive. These serve as orientations for everyday work with artificial intelligence. Universally applicable standards, rules, and laws do not yet exist. However, this could change in the near future.

The European Union (EU) has been discussing a draft law on the regulation of artificial intelligence for some time. Known as the AI Act, this draft has the potential to be a gamechanger for the global AI industry. This is because it is not only European companies that are targeted by this draft law. All companies offering AI systems on the European market, whose AI-generated output is used within the EU, or operate AI systems for internal use within the EU would be affected. The requirements that an AI system must meet depend on its application.

Recruiting algorithms are likely to be classified as high-risk AI. Accordingly, companies would have to fulfill comprehensive requirements during the development, publication, and operation of the AI solution. Among other things, companies are required to comply with data quality standards, prepare technical documentation, and establish risk management. Violations may result in heavy fines of up to 6% of global annual sales. Therefore, companies should already start dealing with the upcoming requirements and their AI algorithms. Explainable AI methods (XAI) can be a useful first step. With their help, black-box algorithms can be better understood, and the transparency of the AI solution can be increased.

Unlocking the Black Box with Explainable AI Methods

XAI methods enable developers to better interpret the concrete decision-making processes of algorithms. This means that it becomes more transparent how an algorithm has formed patterns and rules and makes decisions. As a result, potential problems such as discrimination in the application process can be discovered and corrected. Thus, XAI not only contributes to greater transparency of AI but also favors its ethical use and thus increases the conformity of an AI with the upcoming AI Act.

Some XAI methods are even model-agnostic, i.e. applicable to any AI algorithm from decision trees to neural networks. The field of research around XAI has grown strongly in recent years, which is why there is now a wide variety of methods. However, our experience shows that there are large differences between different methods in terms of the reliability and meaningfulness of their results. Furthermore, not all methods are equally suitable for robust application in practice and for gaining the trust of external stakeholders. Therefore, we have identified our top 3 methods based on the following criteria for this blog post:

Is the method model agnostic, i.e. does it work for all types of AI models?
Does the method provide global results, i.e. does it say anything about the model as a whole?
How meaningful are the resulting explanations?
How good is the theoretical foundation of the method?
Can malicious actors manipulate the results or are they trustworthy?

Our Top 3 XAI Methods at a Glance

Using the above criteria, we selected three widely used and proven methods that are worth diving a bit deeper into: Permutation Feature Importance (PFI), SHAP Feature Importance, and Accumulated Local Effects (ALE). In the following, we explain how each of these methods work and what they are used for. We also discuss their advantages and disadvantages and illustrate their application using the example of a recruiting AI.

Efficiently Identify Influencial Variables with Permutation Feature Importance

The goal of Permutation Feature Importance (PFI) is to find out which variables in the data set are particularly crucial for the model to make accurate predictions. In the case of the recruiting example, PFI analysis can shed light on what information the model relies on to make its decision. For example, if gender emerges as an influential factor here, it can alert the developers to potential bias in the model. In the same way, a PFI analysis creates transparency for external users and regulators. Two things are needed to compute PFI:

An accuracy metric such as the error rate (proportion of incorrect predictions out of all predictions).
A test data set that can be used to determine accuracy.

In the test data set, one variable after the other is concealed from the model by adding random noise. Then, the accuracy of the model is determined over the transformed test dataset. From there, we conclude that those variables whose concealment affects model accuracy the most are particularly important. Once all variables are analyzed and sorted, we obtain a visualization like Figure 1. Using our artificially generated sample data set, we can derive the following: Work experience did not play a major role in the model, but ratings from the interview were influencial.

Figure 1 – Permutation Feature Importance using the example of a recruiting AI (data artificially generated).

A great strength of PFI is that it follows a clear mathematical logic. The correctness of its explanation can be proven by statistical considerations. Furthermore, there are hardly any manipulable parameters in the algorithm with which the results could be deliberately distorted. This makes PFI particularly suitable for gaining the trust of external observers. Finally, the computation of PFI is very resource efficient compared to other XAI methods.

One weakness of PFI is that it can provide misleading explanations under some circumstances. If a variable is assigned a low PFI value, it does not always mean that the variable is unimportant to the issue. For example, if the bachelor’s degree grade has a low PFI value, this may simply be because the model can simply look at the master’s degree grade instead since they are usually similar. Such correlated variables can complicate the interpretation of the results. Nonetheless, PFI is an efficient and useful method for creating transparency in black-box models.

Strengths	Weaknesses
Little room for malicious manipulation of results	Does not consider interactions between variables
Efficient computation

Uncover Complex Relationships with SHAP Feature Importance

SHAP Feature Importance is a method for explaining black box models based on game theory. The goal is to quantify the contribution of each variable to the prediction of the model. As such, it closely resembles Permutation Feature Importance at first glance. However, unlike PFI, SHAP Feature Importance provides results that can account for complex relationships between multiple variables.

SHAP is based on a concept from game theory: Shapley values. Shapley values are a fairness criterion that assigns a weight to each variable that corresponds to its contribution to the outcome. This is analogous to a team sport, where the winning prize is divided fairly among all players, according to their contribution to the victory. With SHAP, we can look at every individual obversation in the data set and analyze what contribution each variable has made to the prediction of the model.

If we now determine the average absolute contribution of a variable across all observations in the data set, we obtain the SHAP Feature Importance. Figure 2 illustrates the results of this analysis. The similarity to the PFI is evident, even though the SHAP Feature Importance only places the rating of the job interview in second place.

Figure 2 – SHAP Feature Importance using the example of a recruiting AI (data artificially generated).

A major advantage of this approach is the ability to account for interactions between variables. By simulating different combinations of variables, it is possible to show how the prediction changes when two or more variables vary together. For example, the final grade of a university degree should always be considered in the context of the field of study and the university. In contrast to the PFI, the SHAP Feature Importance takes this into account. Also, Shapley Values, once calculated, are the basis of a wide range of other useful XAI methods.

However, one weakness of the method is that it is more computationally expensive than PFI. Efficient implementations are available only for certain types of AI algorithms like decision trees or random forests. Therefore, it is important to carefully consider whether a given problem requires a SHAP Feature Importance analysis or whether PFI is sufficient.

Strengths	Weaknesses
Little room for malicious manipulation of results	Calculation is computationally expensive
Considers complex interactions between variables

Focus in on Specific Variables with Accumulated Local Effects

Accumulated Local Effects (ALE) is a further development of the commonly used Partial Dependence Plots (PDP). Both methods aim at simulating the influence of a certain variable on the prediction of the model. This can be used to answer questions such as “Does the chance of getting a management position increase with work experience?” or “Does it make a difference if I have a 1.9 or a 2.0 on my degree certificate?”. Therefore, unlike the previous two methods, ALE makes a statement about the model’s decision-making, not about the relevance of certain variables.

In the simplest case, the PDP, a sample of observations is selected and used to simulate what effect, for example, an isolated increase in work experience would have on the model prediction. Isolated means that none of the other variables are changed in the process. The average of these individual effects over the entire sample can then be visualized (Figure 3, above). Unfortunately, PDP’s results are not particularly meaningful when variables are correlated. For example, let us look at university degree grades. PDP simulates all possible combinations of grades in bachelor’s and master’s programs. Unfortunately, this results in cases that rarely occur in the real world, e.g., an excellent bachelor’s degree and a terrible master’s degree. The PDP has no sense for unreaslistic cases, and the results may suffer accordingly.

ALE analysis, on the other hand, attempts to solve this problem by using a more realistic simulation that adequately represents the relationships between variables. Here, the variable under consideration, e.g., bachelor’s grade, is divided into several sections (e.g., 6.0-5.1, 5.0-4.1, 4.0-3.1, 3.0-2.1, and 2.0-1.0). Now, the simulation of the bachelor’s grade increase is performed only for individuals in the respective grade group. This prevents unrealistic combinations from being included in the analysis. An example of an ALE plot can be found in Figure 3 (below). Here, we can see that ALE identifies a negative impact of work experience on the chance of employment, which PDP was unable to find. Is this behavior of the AI desirable? For example, does the company want to hire young talent in particular? Or is there perhaps an unwanted age bias behind it? In both cases, the ALE plot helps to create transparency and to identify undesirable behavior.

Figure 3- Partial Dependence Plot and Accumulated Local Effects using a Recruiting AI as an example (data artificially generated).

In summary, ALE is a suitable method to gain insight into the influence of a certain variable on the model prediction. This creates transparency for users and even helps to identify and fix unwanted effects and biases. A disadvantage of the method is that ALE can only analyze one or two variables together in the same plot, meaningfully. Thus, to understand the influence of all variables, multiple ALE plots must be generated, which makes the analysis less compact than PFI or a SHAP Feature Importance.

Strengths	Weaknesses
Considers complex interactions between variables	Only one or two variables can be analyzed in one ALE plot
Little room for malicious manipulation of results

Build Trust with Explainable AI Methods

In this post, we presented three Explainable AI methods that can help make algorithms more transparent and interpretable. This also favors meeting the requirements of the upcoming AI Act. Even though it has not yet been passed, we recommend to start working on creating transparency and traceability for AI models based on the draft law as soon as possible. Many Data Scientists have little experience in this field and need further training and time to familiarize with XAI concepts before they can identify relevant algorithms and implement effective solutions. Therefore, it makes sense to familiarize yourself with our recommended methods preemptively.

With Permutation Feature Importance (PFI) and SHAP Feature Importance, we demonstrated two techniques to determine the relevance of certain variables to the prediction of the model. In summary, SHAP Feature Importance is a powerful method for explaining black-box models that considers the interactions between variables. PFI, on the other hand, is easier to implement but less powerful for correlated data. Which method is most appropriate in a particular case depends on the specific requirements.

We also introduced Accumulated Local Effects (ALE), a technique that can analyze and visualize exactly how an AI responds to changes in a specific variable. The combination of one of the two feature importance methods with ALE plots for selected variables is particularly promising. This can provide a theoretically sound and easily interpretable overview of the model – whether it is a decision tree or a deep neural network.

The application of Explainable AI is a worthwhile investment – not only to build internal and external trust in one’s own AI solutions. Rather, we expect that the skillful use of interpretation-enhancing methods can help avoid impending fines due to the requirements of the AI Act, prevents legal consequences, and protects those affected from harm – as in the case of incomprehensible recruiting software.
Our free AI Act Quick Check helps you assess whether any of your AI systems could be affected by the AI Act: https://www.statworx.com/en/ai-act-tool/