Finding a suitable XAI method from the many XAI possibilities for a specific use case is a non-trivial task. There are recommendation algorithms for recommending XAI methods. However, these are often based solely on the underlying ML model's technical characteristics and do not consider the needs of the target group. Also, these systems often recommend only of-the-shelf XAI methods, frequently failing to achieve the desired explanatory goal. We therefore introduce an automated recommendation framework that tackles both of these problems. On the one hand, we created a low-threshold process in which the needs of the target group can be captured in natural language. On the other hand, we recommend XAI Solutions that include both a suitable XAI method and actionable human-centered design guidelines, which describe how the explanation should be adjusted to be useful for the target group. Our recommendation framework consists of a customized GPT that ofers suitable XAI Solutions based on the given design principles and an XAI database. We evaluate our recommendation framework in two real-world scenarios. The evaluation shows that it can generate human-centered XAI solutions that meet the needs of the target group.
The advent of accessible AI technology presents a significant opportunity for individuals without a technical background to realize their ideas for AI systems. While this democratization of AI signifies a beneficial advancement, as it fosters the creation of new interdisciplinary AI systems incorporating a diverse range of ideas, it also introduces risks. At the same time, it emphasizes the need for explainable and transparent AI. This underlines the central role of explainable AI (XAI) in ensuring transparency and control.
Despite the potential benefits of XAI, there are two challenges for applying XAI methods. On the one hand, it is very dificult to select the most suitable method for a specific use case from the large and confusing range of XAI methods. On the other hand, of-the-shelf XAI methods are usually not easily applicable, as explanations for AI systems usually have to be generated in a very context-specific way and the explanations have to be adapted to the specific target group Multimodal, Afective and Interactive eXplainable AI Workshop, 27th European Conference on Artificial Intelligence ⋆You can use this document as the template for preparing your publication. We recommend using the latest version of the ceurart style. * Corresponding author. $ nils.breuer@gt-arc.com (N. O. Breuer)
© 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
to be interpretable. Consequently, the laudable intentions to make the AI system as transparent
and explainable as possible usually fail because an unsuitable XAI method was selected or the
target group does not understand the explanation, as the helpfulness in specific use cases is
questionable [
Our analysis reveals two problem areas in the practical application of XAI methods: first the
selection of an XAI method and second the adaptation of the explanation to a specific target
group. There are solutions for both problem areas. Recommendation algorithms exist to support
the search process for a suitable XAI method. These recommendations are mostly based on
the technical details of the ML model and not on the audience to whom the explanation is
directed [
In this work, we aim to connect and address these two problems with a new human-centered recommendation framework for XAI Solutions. The objective of our recommendation framework is twofold. First, to enable practitioners without extensive experience in XAI to easily access a suitable XAI Solution. This requires a low-threshold and interactive process to gather relevant information. For this, we implemented a natural language-based procedure to collect information about the target group, i.e., the explainees, and also the technical characteristics of the AI model to be explained. Second, we want to recommend XAI Solutions which we define not only as an equivalent for an of-the-shelf XAI method but rather we use the information about the target group to generate a combination of a suitable XAI method and actionable human-centered design guidelines that are based on HCXAI concepts to make the explanation as understandable as possible for the target group. For this, we formulate human-centered design principles based on theories of HCXAI and social sciences. Additionally, we build an XAI database with the most relevant XAI methods.
We create a custom GPT for the recommendation process, using the options provided by OpenAI’s ChatGPT platform. The custom GPT uses knowledge from our human-centered design principles and knowledge from our XAI database to recommend a custom XAI Solution according to our definition.
In summary, our work provides the following contributions:
• We formulate human-centered design principles that are based on socio-cognitive theory
and HCXAI [
Solutions for a specific target group.
Many explanations of XAI approaches currently rely on an algorithm-centric perspective and
are therefore based on the intuitions and explanatory objectives of XAI researchers. These
perspectives and explanatory objectives diverge significantly from the requirements of
laypersonfriendly explanations [
The idea that many HCXAI scientists pursue is that explanations will be more understandable
and useful if they correspond to the social and cognitive processes of human beings. This
humancentered perspective also builds on the article of Miller [
HCXAI has developed many theories and principles to date, as can be seen from the sheer
number of recently published articles [
These theories and principles have already been put into practice in some studies [
The problem of selecting a suitable XAI method is ubiquitous in the literature. The usual
approach is to use one of the many available XAI libraries [
Based on interviews with data scientist Retzlaf et al. [
One of the few works that aim to automate the recommendation process is the AutoXAI
framework by Cugny et al. [
Studies that focus more on the needs of the target group such as [
In this section, we delve deeper into the individual components of our recommendation framework. First, we explain our human-centered design principles, then discuss how we gather information about the target group of the XAI solution, how we map this information to XAI methods, and finally, how the individual components are processed through our recommendation procedure. Figure 1 shows an overview of all the components of our recommendation framework.
As already described in the introduction, existing XAI recommendation algorithms [
From the insights of human-centered XAI research [
• Communication. Social science theories indicate that explaining is a form of social
dialogue [
The full design principles for all categories can be found in Appendix A. These design principles form one of the knowledge files (see Figure 1, blue box) that the recommendation GPT uses to generate the actionable design guidelines of the XAI Solution.
As can be seen from Figure 1, the recommendation GPT receives information about the background knowledge and the explanatory needs of the target group as input.
Background Knowledge In many studies where a target group is characterized to create
more personalized explanations, strict stakeholder groups are defined, each with specific needs
for an explanation [
Explanatory Needs Further important information about the target group includes the needs
they have for an explanation. Explanatory needs can be divided into tasks that the target group
wants to solve with the explanation and goals that the target group wants to achieve with
an explanation [
In addition to the actionable human-centered design guidelines, XAI methods are also part of our XAI Solution. For this purpose, we created an XAI database with common XAI methods. The database defines for each method technical characteristics with which the method is compatible. Beyond that, we also match goals and tasks to the XAI methods so that the explanatory needs of the target group can be used to find a suitable XAI method. For example, the above-mentioned task (What features are most important for the ML model?) is assigned to a feature attribution method.
Due to the high context specificity, the individual requirements of the target group, and also the many diferent types of ML models, it is very challenging to create a structured selection process of XAI methods and human-centered design principles. Many of our attempts with structural databases failed because we always identified a new use case with a new alternative ML model in combination with specific requirements of a target group for which no suitable XAI method in combination with well-founded design guidelines could be found. This is why we have opted for a more open approach. This approach is reflected in the way the information is collected, namely in the form of guided free text.
To match the freely formulated explanatory needs and background knowledge of the target
group with actionable human-centered design guidelines and a suitable XAI method from
our XAI database, we utilize the remarkable association possibilities and few-shot learning
capabilities of OpenAI’s large language model GPT-4 [
We evaluate the capabilities of our recommendation GPT with two use cases. As there is
still no standardized methodology for evaluating the output of LLM, we have evaluated the
XAI solutions of our recommendation GPT in three diferent ways. Human evaluation is still
the gold standard for assessing the quality of LLM output. Because conducting an extensive
human-evaluation study is time-consuming we limited ourselves to a qualitative analysis of the
generated XAI solutions. In addition, we use two automated evaluation metrics to assess the
quality of the design guidelines of the XAI solutions. An embedding-based metric for semantic
textual similarity called SemScore [
The first scenario that we analyzed is that of a senior doctor working in the field of oncology. Recently, the hospital where she works decided to install a new type of AI algorithm to help detect early breast cancer. However, there are concerns among doctors as they have heard of other medical algorithms that have also produced negative aspects. The doctors would therefore like to have an explanation for the algorithm’s decision so that they can evaluate whether the explanation is flawed and at the same time create confidence in working with the algorithm. The whole information retrieval process can be viewed in the Appendix D.1 in addition to the XAI Solution provided by our recommendation GPT. In Figure 2 an excerpt of the information retrieval process for the formal knowledge of the target group is visualized.
In the second scenario, a group of ML engineers is searching for an explanation for an ML algorithm with which extreme weather events can be predicted based on climate data. The problem is that the engineers do not have any formal and instrumental knowledge about the weather data. Therefore, they want to use the explanation to improve the model and also understand the climate data in more detail. Again the whole fictive interaction and the XAI Solution can be found in Appendix D.2. In Figure 3 we visualized the information retrieval process of the instrumental knowledge and the explanatory needs of the target group.
The XAI Solution (Appendix D.1, Figure 4) shows that for each of the four explanatory categories, the recommender GPT gives an actionable design recommendation. The recommendation follows our definitions and design principles which we extracted from HCXAI research. For example, the recommendation is to use clear and concise language that integrates medical terms for the explanation. This corresponds to our rationale for using language as a supporting medium for visual explanations. This shows that the recommender GPT uses the provided knowledge file to generate the design recommendation. The selected XAI methods also meet the needs of the target group. Counterfactual explanation methods are especially often used to create a causal understanding of an ML model. Since the goal of the explanation for the target group is to evaluate how trustworthy the algorithm is, a counterfactual explanation fits well, as it allows decision characteristics to be recognized. These characteristics can then be compared with the physicians’ medical expertise to decide whether the prediction is trustworthy or not. From a technical perspective, both recommended XAI methods can also be used for the ML model.
As for use case 1 the XAI Solution (Appendix D.2, Figure 5) shows some favorable properties for use case 2 that are based on the needs and goals of the target group. For example, one goal is to understand how the model uses climate data to make its prediction. The XAI Solution, therefore, recommends adding an interactive component to the explanation where the engineers can construct diferent scenarios and understand how diferent climate features impact the prediction of the ML model. Additionally, the XAI Solution suggests that the explanation can use highly specific technical language and statistical concepts because the target group consists of ML engineers with several years of working experience. Another objective of the target group is that they want to know which features are most relevant for the ML model, hence the XAI methods recommended in the XAI Solution are suitable, as two of them are feature attribution methods.
We also use two automated evaluation metrics so that the evaluation does not depend solely on
our, possibly biased, perception. SemScore [
XXXXXXXMXetrXicX
Use Case Oncology AI Extreme Weather AI
SemScore
In this study, we embarked on developing a recommendation framework that finds and suggests
human-centered XAI Solutions for individual use cases. For that, we first had to introduce our
notion of XAI Solution which is a combination of conventional XAI methods and actionable
human-centered design guidelines with which these methods can be tailored to a specific target
group. To recommend the design guidelines we formulated human-centered design principles
based on research, drawing on studies from human-centered XAI [
Because a human-centered XAI Solution is highly individual and context-specific, it is almost
impossible to implement a hard-coded and structured approach with database queries or decision
trees. This conclusion has also been reached by numerous predecessors [
We evaluated the results of the recommendation framework for two use cases with three evaluation metrics. The qualitative analysis of the XAI solutions shows that the recommender GPT is capable of using the information about the target group to formulate actionable human-centered design guidelines that correspond to the design principles we formulated. The recommendation framework grasps the important parts of the target group information and matches it to the design principles, for example, for the target group ML engineer it recommends that statistical concepts and machine learning concepts should be used for the explanation. In contrast, for the target group doctor, it is recommended to avoid technical terms of machine learning but rather to convey the explanation in medical jargon. Furthermore, if possible, the explanation should not be centered around ML theories, as the information is not helpful and rather counterproductive for them. The positive direction suggested by our qualitative analysis is also backed by the quantitative evaluation metrics. We focused the automated evaluation metrics on the criteria of semantic similarity. The rationale behind that choice is that we want to assess whether the generated human-centered design guidelines semantically follow our design principles, i.e. have the same meaning. The evaluation with both metrics shows positive results which means that even though both evaluation techniques are based on completely diferent methodologies our recommendation GPT produces trustworthy actionable design guidelines which are in agreement with our design principles. Furthermore, we argue that the framework pays attention to the needs of the target group when selecting a suitable XAI method. In both of the use cases, the recommender provides suitable help on which XAI method should be used to achieve the explanatory goal. For example, the group of ML engineer wants to know which features are most important for the ML model to use this information to improve the ML model. The recommendation framework rightfully suggests using feature attribution methods that can provide precisely this information.
Another distinguishing feature of our recommendation framework is that it is language-based.
This opens up the possibility for people who are not necessarily familiar with AI to describe
their needs in their own words. Natural language enables a whole new level of customization
and users can describe their use cases much more precisely than approaches that only ofer a
limited choice for specific use cases [
Our recommendation framework thus addresses both issues introduced in the introduction regarding XAI. It assists in selecting an appropriate XAI method, considering not only the technical characteristics of the ML model but also selecting the method based on the knowledge level and explanation goals of the target group. Furthermore, the recommendation framework provides recommendations and guidelines based on well-established theories of HCXAI, on how explanations should be tailored to maximize understandability for the target group. To the best of our knowledge, our approach is the first to integrate these two issues and propose a solution. Furthermore, unlike any previous methods, our approach ofers unprecedented individuality in the selection and deployment of XAI methods.
Of course, we are aware of the risks and limitations associated with the use of large language
models [
Unfortunately, due to the basic structure and proprietary nature of most large language models, it cannot be completely ruled out that fatal hallucinations will be generated despite extensive evaluation. However, we believe that the task to be solved involves such a high degree of complexity and that a fundamental feature of the framework is the focus on linguistic exchange that only an LLM is capable of producing satisfactory results.
In this work, we build a human-centered recommendation framework for XAI Solutions for specific use cases. Our XAI Solutions consist of a suitable XAI method and actionable humancentered design guidelines with which the explanation can be tailored for the target group for better understanding. For this, we build a customized GPT that uses well-defined humancentered design principles and an XAI database to generate the XAI Solution. The evaluation of the recommendation framework on two use cases shows that practitioners can overcome two long-lasting problems in the applications of XAI: first finding a good XAI method and second making the explanation understandable to the target group.
There are numerous possibilities for expanding the framework in future work. On the one hand, we aim to expand the human-centered design principles and incorporate even more insights from interdisciplinary sciences. On the other hand, the XAI database can be endlessly expanded with ever-new XAI methods. To make the recommender GPT more robust with larger datasets, methods like retrieval-augmented generation (RAG) could be employed.
This work was conducted as part of the Go-KI project (Ofenes Innovationslabor KI zur Förderung gemeinwohlorientierter KI-Anwendungen), funded by the German Federal Ministry of Labour and Social Afairs (BMAS) under the funding reference number DKI.00.00032.21.
This is the content of our knowledge file which the recommender GPT uses to build the XAI Solution. The knowledge file contains design principles for four explanatory characteristics. COMMUNICATION From a psychological and philosophical perspective, explanations are a form of social dialogue. Explanations based solely on visualizations can be hard to interpret. In addition to just using images and tables, the visual explanation should also be described verbally, like with text. Moreover, each explanation should be accompanied by some sort of interpretation guide. This is especially helpful for people who don’t have a high level of formal, instrumental, and personal knowledge about AI. For AI experts with significant knowledge, the verbal component isn’t as crucial, but an interpretation guide still improves understanding. INTERACTIVITY Simply presenting raw visualizations like saliency maps and tabular explanations does not provide complete understanding. From a social science perspective, explanations are often viewed as an interactive process. Thus, explanations for a machine learning model should also be interpreted and structured as an interactive process. This means there should be a back-and-forth between the explaining medium and the person receiving the explanation. Through this process, most audiences can gain a reliable mental model of the behavior of an ML model.
SELECTIVITY / COMPLEXITY Traditional explanations like saliency maps or tabular visualizations are often cognitively demanding and can overwhelm people who don’t have significant formal, instrumental, or personal knowledge because they are too complex. Social sciences have shown that explanations are selective, meaning the explainer chooses only the most important and relevant causes to build an explanation. To apply this to explainable AI, a way should be found to ask for the needs of the explanation consumer before generating the explanation, allowing the most critical parts to be selected. For example, in feature-importance methods, not all features should be displayed, just those that are most important or of interest to the consumer. Similarly, saliency maps can be simplified to make them less complex. PERSONALIZABILITY Studies have shown that when the recipient of the explanation can customize it to their knowledge level, mental model, and preferences, it positively impacts understanding. These personalizations could include adjusting the complexity or presentation style. However, this should be available only to those with a higher level of AI knowledge. Simpler personalizations, like changing the color scheme or size, could be ofered to those with less technical expertise, giving them opportunities to engage with the explanation and better understand it.
We retrieve the background knowledge and expertise the same way as the explainee needs. We let the user of the framework describe the knowledge of the target group themselves and just give initial ideas on what to describe.
First, we ask if the target group can be categorized into specific stakeholder groups, for example: • Afected: People who do not actively use the system but are afected by its decisions. • Users: People who actively use the system. • Developers: People who implement and build the system • Owners: People who own the system but not necessarily develop or use it. • Validators: People who have a supervisory function.
Second, questions about the formal knowledge of the target group: • What theoretical/formal knowledge do they have about AI and ML learning, e.g. do they know the mathematical background of the ML algorithm? • What theoretical/formal knowledge do they have about the domain in which the ML model is applied? • What is the target group’s level of knowledge about similar types of ML algorithms from the public or from training programs? Third, questions about the instrumental knowledge of the target group • What practical knowledge about programming and ML learning does the target group have? • Does the target group have practical experience in the domain? • Does the target group use a lot of AI systems in general?
First, the user of the framework can name questions that the target group wants an answer for. We ask the question: What type of question does the target group of the explanation want to be answered? and provide some examples of what we mean by the question. • How reliable is the prediction of the ML model? • Is the prediction of my ML model faulty/discriminatory/random? • What information and features does the ML model use to generate the prediction? • Which factors are particularly important for the output of the ML model? • What are the limitations of the ML model? Second the user of the framework can name goals that the target group wants to achieve with the explanation. Again we first ask the question: What goals should be achieved by the explanation? and then provide some examples for inspiration.
• Error detection and improvement of the ML model • I want to ensure that the ML model complies with regulations and laws • I want to understand how the ML model can be used for specific tasks • I want to justify and explain the decisions of the ML model with explanations • I want to know how the data is utilized • I want to learn about the knowledge domain in which the model is deployed • I want to learn how to modify the decisions of the ML model • I want to create trust through explanation so that users of the ML model have a positive experience when using it
The following section is the instruction prompt for the recommender GPT in the OpenAI ChatGPT platform. The JSON files refer to the design principles 3.1 and the XAI database 3.3: This GPT is an assistant that gives human-centered XAI recommendations. The recommendation consists of human-centered design principles which are based on the HCDesignPrinciples.json file and on XAI Methods which should be from the methods.json file.
The recommendation for human-centered design principles consists of recommendations in four characteristics. The recommendation says how explanations should be tailored to a target group in these four characteristics. The characteristics are communication, interactivity, selectivity, and customizability. Guidelines for all four characteristics are discussed in the knowledge file (HCDesignPrinciples.json). Based on these guidelines, recommendations for the four characteristics will be made for a specific target group.
The GPT receives answers to questions that describe the knowledge of the target group and also information about what the explanatory goal of the target group is. Based on this information the GPT matches the guidelines in the knowledge file (HCDesignPrinciples.json) to the information of the target group and makes recommendations for all four characteristics on how an explanation should be adjusted so that it is understandable to the target group.
Q: For whom are you seeking an explanation for? Describe the target group. Can the target group be assigned to one of the following categories (Afected, User, Developers, Owners, Validators)? A: I am seeking an explanation for doctors at a hospital who are the users of an ML algorithm. Q: We will now ask you more about the knowledge of the target group. First, we want to know about the formal knowledge of the target group. What theoretical/formal knowledge do they have about AI and ML learning, e.g., do they know the mathematical background of the ML algorithm? A: The doctors in the hospital do not know anything about AI and ML. Some of them have a solid mathematical background but in general, they do not know how the ML algorithm works in theory.
Q: What theoretical/formal knowledge do they have about the domain in which the ML model is applied to? A: The doctors who use the algorithm are well-trained doctors in the domain of oncology specifically breast cancer.
Q: What is the target group’s level of knowledge about similar types of ML algorithms from the public or from training programs? A: Before the algorithm was used, all users attended an introduction and information event in which the use of the algorithm was explained.
Q: Now we will ask about the instrumental knowledge of the target group. What practical knowledge about programming and ML learning does the target group have? A: The doctors using the ML system do not have any knowledge about practical programming or ML learning.
Q: Does the target group have practical experience in the domain? A: The target group works as doctors and is experienced in practice.
Q: Does the target group use a lot of AI systems in general? A: All users have an average amount of contact with AI in general.
Q: What type of question does the target group of the explanation want to be answered? (E.g., How faithful is the prediction of the algorithm? Which features are most important for the algorithm?) A: The target group wants to know if the prediction is incorrect and also how trustworthy the prediction of the algorithm is.
Q: What goals should be achieved by the explanation? A: The goal is to use the explanation to build trust among the doctors using the algorithm but also to justify the decision that we make based on the algorithm’s prediction.
Q: For whom are you seeking an explanation for? Describe the target group. Can the target group be assigned to one of the following categories (Afected, User, Developers, Owners, Validators)? A: The target group is ML Engineers working at an AI software company that is implementing an ML algorithm.
Communication The explanation should employ clear, concise language that integrates medical terminology familiar to oncologists with basic statistical concepts to explain the algorithm's outputs. Use of visual aids like graphs or charts to show statistical reliability (like confidence intervals or prediction probabilities) can be helpful. Since the doctors have attended an introductory event, the explanation can build on the foundational knowledge they already possess about the algorithm’s intended use, focusing on practical application rather than theoretical aspects Selectivity / Complexity The explanation should selectively focus on aspects of the algorithm that directly relate to clinical decision-making in oncology.
This might include an emphasis on the most significant features the algorithm considers when making a prediction, such as tumor size, type, and location, and how these are weighted within the model. Presenting simplified case studies where the algorithm's predictions were particularly successful, as wel as scenarios where it was less accurate, can il ustrate its performance spectrum in a clinical y relevant manner.
Counterfactual Explanations: This method would al ow doctors to see how altering certain input variables (like changing a measurement or condition slightly) could lead to different predictions. This is particularly useful for understanding edge cases or close-cal decisions made by the algorithm.
Feature Importance: Methods like SHAP (SHapley Additive exPlanations) can elucidate which features are most influential in the algorithm’s predictions. This aligns wel with a need to understand and trust the predictive factors, especial y in complex cases of breast cancer diagnosis.
These explanations, tailored to the needs and backgrounds of the doctors, aim to enhance their trust in the AI system by making its operations transparent and directly relevant to their clinical practice. This approach should also aid in justifying clinical decisions influenced by the algorithm's predictions, ensuring that the tool is a valuable asset to the oncological team. the ML algorithm? and ML learning algorithms. model is applied to? general. We only use the data. the public or from training programs? We will now ask you more about the knowledge of the target group. First, we want to know about the formal knowledge of the target group. What theoretical/formal knowledge do they have about AI and ML learning, e.g., do they know the mathematical background of We are all ML Engineers and have extensive knowledge about the theory behind most AI Q: What theoretical/formal knowledge do they have about the domain in which the
ML A: The algorithm that we are implementing is concerned with predicting extreme weather events from climate data. We do not have knowledge about meteorology and climate in Q: What is the target group’s level of knowledge about similar types of ML algorithms from We researched similar algorithms online that also use climate data to predict extreme weather events.
Now
we will ask about the instrumental knowledge of the target group. What practical knowledge about programming and ML learning does the target group have?
Communication Given the target group’s extensive knowledge in ML, the explanations can use technical language, incorporating advanced statistical and machine learning concepts. However, since the engineers are not familiar with meteorology, it’s crucial to clearly define any meteorological terms and concepts used in the model. Visual aids like advanced data visualizations that map climate data inputs to prediction outputs can be beneficial.
Documentation or tooltips that provide brief explanations of meteorological terms and their relevance to the model’s predictions can enhance understanding.
Interactivity An interactive explanation interface would be highly beneficial for this group. It could al ow engineers to manipulate input data points and observe how these changes affect the model’s predictions. Tools that enable them to test model behavior under various synthetic scenarios they construct can deepen their understanding of how different climate features influence predictions. This interactive process facilitates a practical exploration of the model's behavior, providing insights that are crucial for both trust and model refinement.
Customizability Providing options to customize the level of detail in the explanations would cater to individual preferences within the group. Some engineers might prefer a high-level overview of model performance metrics, while others might delve into the nuances of algorithmic decisions and data usage.
Al owing engineers to choose the complexity of the information presented or to toggle between a basic and advanced view of explanations could enhance usability and engagement. are.
We are all experienced ML engineers with several years of working experience. Q: Does the target group have practical experience in the domain? A: No, we do not have any knowledge about meteorology.
Q: Does the target group use a lot of AI systems in general?
We are very proficient with AI systems and frequently use them in our daily lives. Q: What type of question does the target group of the explanation want to be answered? (E.g., How faithful is the prediction of the algorithm? Which features are most important We want to learn the behavior of the model and want to know which features are most important for the ML model. We also want to know what the limitations of the ML model Q: What goals should be achieved by the explanation? A:
We want to improve the ML model by knowing the behavior of the model and additionally, we want to know how the data is used by the ML model.