XAI or eXplainable Artificial Intelligence is one of the most remarkable fields in computer science and Artificial Intelligence (AI) due to its application in many critical domains such as medicine, defence or industry [? ]. Understanding the reason behind the behaviour of an AI can make a difference in the effectiveness of that model. A central challenge to provisioning XAI is the variety of situations where we need explanations, regarding different domains, users, goals or problems to solve [1]. This exacerbates the decision making task of selecting the right explanation approaches to apply. The iSee (Intelligent Sharing of Explanation Experience by Users for Users) project 1 was proposed with the objective of tackling this challenge. The final goal of the iSee project is to build a platform where users can share and reuse explanation experiences, a concept that comprises all the elements involved in the explanation process: input data, user's goals, human-computer interaction, AI model, explanation algorithms, explanation performance, etc. Here, Case-Based Reasoning (CBR) appears as a natural approach to reuse explanation experiences developed by other users for different models and domains.
To support this CBR process, one of the main results we want to achieve in iSee is the formalisation of a conceptual model that defines an explanation experience. This model has been instantiated into the iSeeOnto ontology, that is able to represent the different conceptual dimensions that describe an explanation experience, mainly: (1) the AI model to explain; (2) the explainer we need to do that and; (3) the user who needs the explanation. Moreover, for each explanation experience, iSeeOnto defines the solution of the problem we want to solve, i.e. the features of the explainer that solves the problem, its implementation, and the methodology to evaluate its performance when deployed into an actual use case (the result of the experience). This way, iSeeOnto provides the conceptual support to describe the explanation experiences as cases, enabling the application of the CBR paradigm to reuse them.
The paper is structured as following. Section 2 describes some literature related to XAI and explanation conceptualisation. Section 3 depicts the current version of the iSeeOnto and the main concepts defined in the ontology. In Section 4, we validate our model, describing explanation systems proposed in the state-of-the-art with our ontology, and building a first approach of the case base we need to build for our iSee platform. Finally, we extract some conclusions from the work done in Section 5.
There is an appreciable amount of surveys, reviews and taxonomies about discovering the concepts related to XAI methods. A recent survey explores XAI methods for deep learners which proposes a taxonomy of three main concepts: scope which identifies the entity explained by the method; methodology which recognises the algorithmic approach to explanation; and usage which refers to the applicability of the method to any or specific methods [2]. In addition to characterising XAI methods, authors of [3] also recognise user needs (i.e. intent) such as effectiveness, transparency, persuasiveness and usefulness and the type of explanations that satisfy these needs. The match between intents and explanations were derived from literature that used either user studies or empirical evaluations. Conceptualisation of XAI by the authors of [4] is another prominent contribution in XAI. It depicts explanations in machine learning and deep learning, while classifying XAI methods with respect to several aspects, such as explanation scope, transparency, data, and visualisation modes, etc.
Our previous work [5], also developed a conceptual model, which extended previous work [3,4] to include new concepts to capture user knowledge. In ExRecOnto ontology we guide the design of explanation systems for recommenders through an ontology-based methodology. When developing iSeeOnto, we further extend our work [5] to capture facets of XAI as a user experience. We also refer to different aspects highlighted in XAI taxonomies found in literature such as user requirements [6] and evaluation of XAI methods [7].
One of the goals of iSee is to develop a CBR system that recommends appropriate XAI methods for AI systems by capturing successful cases of building XAI systems (i.e. explanation . CBR systems which make use of ontologies is a common approach seen in literature. We view the usage of iSeeOnto in two main knowledge sources for the CBR system [8]. Firstly, an ontology can be used as the vocabulary that represents the cases, in a similar way to the work in [9] and [10]. Secondly, iSeeOnto can be viewed as similarity knowledge for retrieval and reuse similar to work in [8] and [11]. In this paper we focus on the first knowledge container and we explore how to represent an XAI user experience as a case using iSeeOnto.
iSeeOnto is an ontology that describes explanation experiences for user-centered XAI. It enables the definition of explanation experiences as cases, consisting of description, solution and result, that can be reused by the iSee CBR engine.
Figure 1 presents the complete view of the iSeeOnto and how it forms the case structure for iSee CBR system. A case describes an explanation experience that captures how an explanation strategy satisfies certain explanation needs. Thus, the ontology defines three main features that describe an explanation experience: (1) the AI model we need to explain; (2) the explainer features we require to explain the AI model and; (3) the user and their explanation requirements.
The solution of an explanation experience formalises how we compose several explanation artefacts to implement a working system that achieves the requirements defined in the description. This way, solutions must capture the combination of explanation components, user evaluation methods, and user-explainer interaction. Workflows are the most common paradigm to implement such needs as they define the specific sequence of steps (or tasks) involved in the process of getting a work done. However, workflows put the user's decision at the centre, not the process itself, and are typically coupled to a particular application instead of being reusable as required by the iSee platform. Instead, we propose the use of Behaviour Trees (BTs) that are mathematical models of plan execution in a modular fashion. They define how a given objective should be achieved in a domain independent way, and in how many different states the information can be during that process. As a result, the process that is described using BTs is independent of the end-user's roles and is modular and reusable, the latter being a major requirement of the CBR adaptation process.
Finally, the UserEvaluation describes how an Explanation can be evaluated, defining the Metric and the Dimension. Next, following subsections describe each one of the previous concepts.
AIModel concept and related features are presented in Figure 2. AIModel utilises an AIMethod which has sub classes like k-NN, SVM and Neural Networks. AIModel is trained on a Dataset which has a DataType (i.e. tabular, image, text, etc.) and characterised by number of features and number of instances. AIModel solves an AITask such as classification, regression or anomaly detection which is also associated with an AITaskGoal. iSeeOnto also characterises how an AIModel is evaluated (not included in Figure 2). The AIModel has AIModelAssessmentResult from when it is evaluated for a AIModelAssessmentDimension such as performance, bias or robustness using a AIModelAssessmentMetric like accuracy, f1-score or recall.
Considering the Explainer, iSeeOnto defines several concepts related to its features as shown in Figure 3. The Explainer utilises an ExplainabilityTechnique that generates an Explanation. An Explanation can be a Feature Importance Explanation, a Contextual Explanation or a Instancebased Explanation, etc. An ExplainabilityTechnique can be applicable only to a certain AI method or an AI task. The ExplainabilityTechnique generates an Explainer which has a representation characterised InformationContentEntity. A few sub classes of InformationContentEntity are visual, text, annotations and charts.
The ExplainabilityTechnique is characterised by many features. Portability indicates if the technique is AI model-agnostic (works for any AI model), AI model-specific (only works for a specific AI model) or model-class specific (only works for a specific type of AI models). ExplainerConcurrentness is ante-hoc when the explanation is generated by the AI model itself; it is considered post-hoc when the explainability technique is independent of the AI model. The ExplainabilityTechnique has an ExplanationScope: local when it explains a single data point or a single prediction; global when it explains the behaviour of the AI model or data as a whole; and cohort when the explanation is related to a subset of predictions or data. An ExplanationTarget is related to the previous concept; it determines the target of the explanation, the predictions, the model or the data.
Regarding the user, iSeeOnto considers concepts depicted in Figure 4. A user has an Intent that is expressed as a UserQuestion. An Intent describes the need for an explanation which can fall under a category such as Transparency, Trust, Effectiveness, Efficiency, Scruitability, Satisfaction, Persuasiveness, Education or Debugging. The UserQuestion can take the form of a How, Why, What, When, Where, or What-if question. The UserQuestionTarget of the UserQuestion points to the aspect of the AI model's behaviour is being questioned by the user (i.e. system recommendation, AI model or data. iSeeOnto also recognises that an Explanation addresses a certain UserQuestion.
Knowledge possessed by the User can take two forms: Domain Knowledge and AI Knowledge both of which can be measured as low, medium or high. The User has some resources identified by TechnicalFacilities such as touch, audio or visual determine which explanation modalities can be presented. For instance, an Explanation that has a interactive presentation is only suitable if the user has a touch or clickable interface. Sub classes of CompositeNode and DecoratorNode concepts carry the standard functionality as intended in generic Behaviour trees. In this paper we highlight few custom nodes for Explanation Strategy representation. The ExplanationMethodNode represents an action node that invokes an API call to a URL (in node property URL) that will fetch an explanation given pre-requisite parameters (also included in node properties). The EvaluationMethodNode will fetch a questionnaire from a URL that will be executed as a sub-tree. The ConditionNode will control the entry to a sub-tree which is also known as memory controlled flow for efficient execution of BTs.
The UserEvaluation describes how an Explanation is evaluated, defining the Metric and the Dimension. An Explanation is annotated by a UserEvaluationResult which is based on a Metric.
To validate the completeness of iSeeOnto regarding its capability to represent an explanation experience properly, we have carried out a validation consisting of the description of thirteen explanation approaches proposed in the literature with the concepts defined in iSeeOnto. We selected these explanation approaches trying to cover a wide variety of features and types. We extracted most of them from the work by Barredo-Arrieta et al. [4]. We also selected some works we used in the validation carried out in our previous paper [5] for the recommender systems domain. Finally, a few of the publications were selected making a systematic search on Google Scholar from December 2021 until February 2022. The keywords used in the search were: "XAI taxonomy", "explanation systems artificial intelligence" and "explanations artificial intelligence" combining these concepts with the ones defined in the taxonomy by Barredo-Arrieta et al. We filtered results for publications not older than ten years. This way, we could represent an amount of different explanation approaches with our ontology, validating its completeness.
Following, we describe every explanation approach that was successfully represented using iSeeOnto and the concepts specified in Section 3. The details of this classification cannot be included in this paper due to space restrictions but are available on GitHub2 . The schema of the concepts used to classify each explanation approach are presented in Figure 7. Below we introduce the explainers we have examined.
PSIE [12]. This paper proposes to recommend movies to groups based on social knowledge.
DisCERN [13]. The goal of this paper is to predict lung cancer risk given clinical data of patients.
BTTelecom [14]. In this paper, authors propose to recommend engineering notes to desk support staff to help on-site engineers.
SciNet [15]. In this work, the objective is to determine the most related documents given a set of keywords. One of the most important features of this approach is that users can interact with the interface to manipulate the keywords and change the search result.
TalkExplorer [16]. This paper proposes to recommend scientific publications based on content and social connections. In this case, an explanation system is incorporated into a conference recommender system. The conferences are grouped in bubbles according to their content. Users can interact with the interface to get more details of the recommendation.
IntGradImage, IntGradRetinopathy and IntGradTextClassification [17]. In this work, the authors propose the explanation technique Integrated Gradients. With this technique, they are able to predict the category of a given image, predict if a given medical image contains diabetic retinopathy and predict the question category based on question text, respectively for each approach.
KimEtAlMethod [18]. The system proposed in this paper makes acceleration or change course decisions in a self-driving car based on video.
iBCM [19]. The main goal of this work is to cluster student assignment submissions to design grading rubric or to compose feedback.
InNoCBR [20]. This publication describes a CBR system with explanations for detecting healthcare associated infections (HCAIs). The system predicts patient's infection based on a clinical, laboratory, and medico administrative based data. [21]. In this work, the SHAP method is proposed to predict patient mortality based on clinical, nutritional and behavioural factors.
From this representation of such a heterogeneous collection of explainers in iSeeOnto, we can support its semantic completeness and extract some conclusions about the features of the explanation systems we have been validated.
Regarding the AIModel and the AITask concepts, we can see that almost half of the approaches explain AI models related to disease prediction using classification tasks. This is also the most common AITaskType: multi-class and binary classification.
Considering the Explainer features, there are several types of explanations, not having a type of Explanation that stands out over others. However, we can observe that some of the explaination approaches for recommender systems are neighbourhood-based explanations. Regarding Portability, we have more post-hoc systems than ante-hoc systems. When the explainer is ante-hoc, the explanation is generated by the AI model itself, so the explainer cannot be decoupled, and therefore it not reusable as the solution for a different problem by a CBR system. According to the InformationContentEntity concept, any entity is the widest option although specific values such as text and image are other options that are also quite common. The ExplanationScope is one of the concepts where we can see uniform results: most of the explainers are local. The same happens with ExplanationTarget, where we can see that most of the explainers has prediction as target.
Taking into account the classification with concepts related to the User, we can determine we have been able to represent a broad variety of users, depending on the domain of application. However, the majority of the UserQuestion of the approaches are Why questions, having only two What/How questions, one Why/How question and one How/What-if question. Then, we can conclude that one of the main users' goals seems to be the understanding of the underlying AIModel. One of the most repeated Intents of the User is Transparency, followed by Trust, which makes sense considering the type of questions the users want to answer about the AI model in our validation. The TechnicalFacilities required by the user are unanimous: in all the explanation experiences studied, users need Screen Display to watch the explanation. Regarding the user knowledge level, most of the users have high level of domain knowledge, and low knowledge level in AI.
Considering the Solution concept, we have to analyse the results from the UserEvaluation. As we can see in the spreadsheet, half of the explanation experiences studied have not carried out an evaluation. There are five approaches evaluated, all of them use questionnaires as the evaluation metric. The dimensions measured are heterogeneous, being Usefulness and Efficiency two of the most common dimensions among the explanation experiences evaluated. Finally, we have to discuss the code availability of the implementation of the solution: only seven of the explanation experiences have the source code available.
Nowadays, trust in AI systems is essential for the acceptance of their predictions by the final user. Therefore, including explanations in intelligent systems is critical to make this type of systems effective.
Fortunately, the XAI community is generating novel explanation methods continuously that address this problem. However, it raises the challenge to find the most suitable explanation strategy for each AI system, being CBR a natural solution for reusing previous explanation experiences. Developing such solution is the goal of the iSee project. As the first step, we have developed the iSeeOnto ontology, which formalises a conceptual model to define and describe explanation strategies and all the elements involved: the AI model to explain, the explainer features and the user requirements.
To check the completeness of iSeeOnto, it has been validated by describing several repre-sentative explanation approaches extracted from the state-of-the-art. The secondary goal of this validation is the collection of an initial case base for the future iSee's CBR platform. From the results of this validation, we can conclude that iSeeOnto is able to classify all of these approaches. Therefore we can propose our ontology as a guideline for designing explanation strategies through an ontology-based methodology. Finally, as future work, we want to extend the validation and the resulting case base with further use cases. In addition, we want to establish the retrieval strategy required to reuse these explanation experiences for different domains, AI models and user's intends.
Supported by the Horizon 2020 Future and Emerging Technologies (FET) programme of the European Union through the iSee project (CHIST-ERA-19-XAI-008, PCI2020-120720-2). Funding in Spain by MCIN/AEI/10.13039/501100011033 and in UK by EPSRC grant number EP/V061755/1.