As a first step towards personalised, interactive XAI systems we developed Glass-Box : a class-contrastive counterfactual explainability system that can be queried with a natural language dialogue (described in Sect. 3.1). It supports a range of “Why?” questions that can be posed either through a voice- or chat-based interface. Building this system and testing it in the wild provided us with invaluable experience and insights, which we now share with the community as they may be useful to anybody attempting to develop and deploy a similar system—Sects. 3.2 and 3.3 discuss interactive explainers desiderata and properties respectively. The feedback that helped us to refine our idea of interactive XAI systems producing personalised explanations (presented in Sect. 3.4) was collected while demonstrating Glass-Box to a diverse audience consisting of both domain experts, approached during the 27th International Joint Conference on Artificial Intelligence (IJCAI 2018), and a lay audience, approached during a local “Research without Borders” festivalFootnote 6 that is open to the public and attended by pupils from local schools. While at the time of presentation our system was limited to class-contrastive counterfactual explanations personalised by (implicitly) choosing data features that the counterfactual statements were conditioned on and provided to the user in natural language, we believe that our observations remain valid beyond this particular XAI technique. We hope to test this assumption in our future work—see Sect. 4 for more details—by employing the remaining four decision tree explainability modalities listed in the introduction, albeit in an XAI system refined based on our experience to date.
Glass-Box has been designed as a piece of hardware built upon the Google AIY (Artificial Intelligence Yourself) Voice KitFootnote 7—a customisable hardware and software platform for development of voice interface-enabled interactive agents. The first prototype of Glass-Box utilised the Amazon Alexa skill Application Programming Interface, however the limitations of this platform at the time (the processing of data had to be deployed to an on-line server and invoked via an API call) have hindered the progress and prompted us to switch to the aforementioned Google AIY Voice device. These recent technological advancements in automated speech-to-text transcription and speech synthesis provided as a service allowed us to utilise an off-the-shelf, voice-enabled, virtual, digital assistant to process explainees’ speech and automatically answer their questions—something that would not have been feasible had we decided to build this component ourselves. We extended the voice-driven user interface with a (textual) chat-based web interface that displays the transcription of the conversation and its history—to improve accessibility of the system, among other things—in addition to allowing the explainee to type in the queries instead of saying them out loud.
To avoid a lengthy and possibly off-putting process of submitting (mock) personal details—i.e., a data point—to be predicted by the underlying Machine Learning model and explained by Glass-Box, we opted for a predefined set of ten data points. Any of them could be selected and input to Glass-Box by scanning a QR code placed on a printed card that also listed details of this fictional individual.
Once a data point is selected, the explainee can alter personal details of this fictional individual by interacting with Glass-Box, e.g., “I am 27 years old, not 45.” Any input to the system is passed to a natural language processing and understanding module built using rasaFootnote 8. Our deployment of the Glass-Box system was based on the UCI German credit data setFootnote 9 (using a subset of its features) for which a decision tree classifier was trained using scikit-learnFootnote 10 . Since the German credit data set has a binary target variable (“good” or “bad” credit score), the class contrast in the counterfactual explanations is implicit. Nevertheless, this could be easily generalised to a multi-class setting by requiring the explainee to explicitly specify the contrast class, taking the second-most likely one or providing one explanation per class. A conceptual design of Glass-Box is presented in Fig. 2.
To facilitate some of the user interactions the data set had to be manually annotated. This process allowed the generation of engaging natural language responses and enabled answering questions related to individual fairness. The latter functionality was achieved by indicating which features (and combinations thereof) should be treated as protected attributes (features), hence had a counterfactual data point conditioned on one of these features been found, Glass-Box would indicate unfair treatment of this individual. This functionality could be invoked by asking “Is the decision fair?” question and further interrogating the resulting counterfactual explanation if one was found. Depending on the explainability and interactiveness requirements expected of the system, other data set annotations may be required. Since annotation is mostly a manual process, creating them can be time- and resource-consuming.
As noted before, the main objective of Glass-Box is to provide the users with personalised explanations whenever they decide to challenge the decision of the underlying Machine Learning model. The explainee can request and interactively customise the resulting counterfactual explanations through a natural language interface with appropriate dialogue cues. This can be done in three different ways by asking the following questions:
“Why?”—a plain counterfactual explanation—the system returns the shortest possible class-contrastive counterfactual;
“Why despite?”—a counterfactual explanation not conditioned on the indicated feature(s)—the system returns a class-contrastive counterfactual that does not use a specified (set of) feature(s) as its condition; and
“Why given?”—a (partially-)fixed counterfactual explanation—the system returns a counterfactual that is conditioned on the specified (set of) feature(s).
By repetitively asking any of the above “Why?” questions the system will enumerate all the possible explanations with the condition set (the features that need to change) increasing in quantity until no more explanations can be found. It is also possible to mix the latter two questions into “Why given ... and despite ... ?”, thereby introducing even stronger restrictions on the counterfactual explanations. In addition to “Why?” questions the explainee can also ask “What if?” In this case it is the user who provides the contrast and wants to learn the classification outcome of this hypothetical data point. This question can be either applied to the selected data point (which is currently being explained) or any of the counterfactual data points offered by the system as an explanation. All of these requirements imposed by the user are processed by a simple logical unit that translates the user requests into constraints applied to the set of features that the counterfactual is allowed and/or required to be conditioned upon. All of these happen through a natural language dialogue, an example of which is depicted in Fig. 3.
The method used to generate counterfactual explanations from the underlying decision tree classifier relies upon a bespoke leaf-to-leaf distance metric. It allows to find leaves of different classes to the one assigned to the selected data point that require the fewest possible changes to this data point in its feature space. One obvious solution to this problem is any neighbouring leaf of a different class; this requires just one feature to be altered. However, there may also exist leaves that are relatively distant in the decision tree structure but also require just one feature value change, for example, when these two decision tree paths do not share many features. This distance metric is computed by representing the tree structure in a binary meta-feature space that is created by extracting all the unique feature partitions from the splits of the decision tree. Finally, an L1-like metric (when a particular feature is present on one branch and absent on the other, this distance component is assumed to be 0) is calculated and minimised to derive a list of counterfactual explanations ordered by their length.
During the development stage and early trials of Glass-Box we identified a collection of desiderata and properties that should be considered when building such systems. Some of these attributes are inspired by relevant literature [20, 26, 34, 44], while others come from our experience gained in the process of building the system, presenting it to various audiences, discussing its properties at different events and collecting feedback about interacting with it. While this and the following sections focus on desiderata for interactive and customisable explanations, we provide an in-depth discussion on this topic for generic explainability systems in our work on “Explainability Fact Sheets” . The relevant subset of these desiderata are summarised in Table 1 as well as collected and discussed below. Section 3.3, on the other hand, examines the properties of interactive explainability systems.
Given the complex nature of such systems, it would be expected that some of these objectives might be at odds with each other, their definition may be “fuzzy”, they might be difficult to operationalise, their “correct” application might depend on the use case, etc. Furthermore, striking the right balance between these desiderata can be challenging. Nevertheless, we argue that considering them while designing interactive explainers will improve the overall quality of the system, help the designers and users understand their strengths and limitations, and make the interaction feel more natural to humans. Furthermore, some of these desired properties can be achieved (and “optimised” for the explainees) by simply allowing user interaction, thereby alleviating the need of explicitly building them into the system. For example, interactive personalisation of the explanations (on-line, with user input) can mean that it does not have to be solved fully algorithmically off-line.
The main advantage of Glass-Box interactiveness is the explainee’s ability to transfer knowledge onto the system—in this particular case various preferences with respect to the desired explanation—thereby personalising the resulting explanation [36, property U10, see Table 1]. In our experience, personalisation can come in many different shapes and forms, some of which are discussed below. By interacting with the system the explainee should be able to adjust the breadth and scope of an explanation [36, property F4]. Given the complexity of the underlying predictive model, the explainee may start by asking for an explanation of a single data point (black-box prediction) and continue the interrogation by generalising it to an explanation of a sub-space of the data space (a cohort explanation) with the final stage entailing the explanation of the entire black-box model. Such a shift in explainee’s interest may require the explainability method to adapt and respond by changing the target of the explanation [36, property F3]. The user may request an explanation of a single data point or a summary of the whole data set (training, test, validation, etc.), an explanation of a predictive model (or its subspace) or any number of its predictions. Furthermore, interactive personalisation of an explanation can increase the overall versatility of such systems as customised explanations may serve different purposes and have different functions [36, property O7]. An appropriately phrased explanation may be used as an evidence that the system is fair—either with respect to a group or an individual depending on the scope and breadth of the explanation—or that it is accountable, which again can be investigated with a varied scope, for example, a “What if?” question uncovering that two seemingly indistinguishable data points yield significantly different class assignment, aka adversarial examples . Importantly, if the explainer is flexible enough and the interaction allows such customisation, however the explanations were designed to serve only one purpose, e.g., transparency, the explainee should be explicitly warned of such limitations to avoid any unintended consequences. For example, the explanations may be counterfactually actionable but they are not causal as they were not derived from a causal model [36, property O8].
Some of the aforementioned principles can be observed in how Glass-Box operates. The contrastive statements about the underlying black-box model can be used to assess its transparency (their main purpose), fairness (disparate treatment via contrastive statements conditioned on protected attributes) and accountability (e.g., answers to “What if?” questions that indicate an unexpected non-monotonic behaviour). The contrastive statements are personalised via user-specified constrains of the conditional part (foil) of the counterfactual explanation and by default are with respect to a single prediction. Cohort-based insights can be retrieved by asking “What if?” questions with regard to counterfactual explanations generated by Glass-Box—Sect. 4 discusses how the scope and the target of our explanations can be broadened to global explanations of the black-box model. Given the wide range of possible explanations and their uses some systems may produce contradictory or competing explanations. Glass-Box is less prone to such issues as the employed explainer is ante-hoc [36, property F7], i.e., predictions and explanations are derived from the same ML model, hence they are always truthful with respect to the predictive model. This means that contradictory explanations are indicative of flaws in the underlying ML model, hence can be very helpful in improving its accountability.
In day-to-day human interactions we are able to communicate effectively and efficiently because we share common background knowledge about the world that surrounds us—a mental model of how to interact with the world and other people . Often, human–machine interactions lack this implicit link making the whole process feel unnatural and frustrating. Therefore, the creators of interactive explainability techniques should strive to make their systems coherent with the explainee’s mental model to mitigate this phenomenon as much as possible [36, property U7]. While this objective may not be achievable in general, modelling a part of the user’s mental model, however small, can make a significant difference. The two main approaches to extracting an explainee’s mental model are interactive querying of the explainee in an iterative dialogue (on-line), or embedding the user’s characteristics and preferences in the data or in the parameters of the explainer (off-line), both of which are discussed in Sect. 2.
For explainability systems this task is possible to some extent as their operation and purpose are limited in scope in contrast to more difficult tasks like developing a generic virtual personal assistant. Designers of such systems should also be aware that many interactions are underlined by implicit assumptions that are embedded in the explainee’s mental model and perceived as mundane, hence not voiced, for example, the context of a follow-up question. However, for human–machine interactions the context and its dynamic changes can be more subtle, which may cause the coherence of the internal state of an explainer and the explainee’s mental model to diverge [36, property U3]. This issue can be partially mitigated by explicitly grounding explanations in a context at certain stages, for example, whenever the context shifts, which will help the users to adapt by updating their mental model and assumptions. Contextfullness will also help the explainee better understand the limitations of the system, e.g., whether an explanation produced for a single prediction can (or must not) be generalised to other (similar) instances: “this explanation can be generalised to other data points that have all of the feature values the same but feature \(x_5\), which can span the \(0.4 \le x_5 < 1.7\) range.”
Regardless of the system’s interactivity, the explanations should be parsimonious—as short as possible but not shorter than necessary—to convey the required information without overwhelming the explainee [36, property U11]. Maintaining a mental model of the user can help to achieve this objective as the system can provide the explainee only with novel explanations—accounting for factors that the user is not familiar with—therefore reducing the amount of information carried by the explanation [36, property U8]. Another two user-centred aspects of an explanation are its complexity and granularity [36, property U9]. The complexity of explanations should be adjusted according to the depth of the technical knowledge expected of the intended audience, and the level of detail chosen appropriately for their intended use. This can either be achieved by design (i.e., incorporated into the explainability technique), be part of the system configuration and parametrisation steps (off-line) or adjusted interactively by the user as part of the explanatory dialogue (on-line). Another aspect of an explanation, which is often expected by humans , is the chronology of factors presented therein: the explainee expects to hear more recent events first [36, property U6]. While this property is data set-specific, the explainee should be given the opportunity to trace the explanation back in time, which can easily be achieved via interaction.
Glass-Box attempts to approximate its users’ mental models by mapping their interests and interaction context (inferred from posed questions) to data features that are used to compose counterfactual explanations. Memorising previous interactions, their sequence and the frequency of features mentioned by the user help to achieve this goal and avoid repeating the same answers—once all of the explanations satisfying given constraints were presented, the system explicitly states this fact. Contextfullness of explanations is based on user interactions and is implicitly preserved for follow-up queries in case of actions that do not alter the context and are initiated by the user—e.g., interrogative dialogue. Whenever the context shifts—e.g., a new personalised explanation is requested by the user or an interaction is initiated by Glass-Box—it is explicitly communicated to the user. Contrastive explanations are inherently succinct, but a lack of parsimony could be observed for some of Glass-Box explanations, which resulted in a long “monologue” delivered by the system. In most of the cases this was caused by the system “deciding” to repeat the personalisation conditions provided by the user to ensure their coherence with the explainee’s mental model.
Glass-Box is capable of producing novel explanations by using features that have not been acknowledged by the user during the interaction. Interestingly, there is a trade-off between novelty of explanations and their coherence with the user’s mental model, which we have not explored when presenting our system but which should be navigated carefully to avoid jeopardising explainee’s trust. Glass-Box was built to explain predictions of the underlying ML model and did not account for possible generalisation of its explanations to other data points (the users were informed about it prior to interacting with the device). However, the explainees can ask “What if?” questions with respect to the counterfactual explanations, e.g., using slight variations of the explained data point, to explicitly check whether their intuition about the broader scope of an explanation holds up. Finally, chronology was not required of Glass-Box explanations as the data set used to train the underlying predictive model does not have any time-annotated features.
In addition to a set of interactive explainability system desiderata, we consider a number of their general properties and requirements that should be considered prior to their development. These are summarised in Table 2 and discussed below.
Assuming that the system is interactive, the communication protocol between the explainee and the explainer should be carefully chosen to support the expected input and deliver the explanations in the most natural way possible. For example, clearly indicating which parts of the explanation can be personalised and the limitations of this process should be disclosed to the user [36, property O3, see Table 2]. The choice of explanatory medium used to convey the explanation is also crucial. Plots, interactive or not, can be very informative, but may not convey the whole story due to the curse of dimensionality and the limitations of the human visual system [36, property O2]. Supporting visualisations with textual description can greatly improve their intelligibility, and vice versa, nevertheless in some cases this approach may be sub-optimal, for example, explaining images using only a natural language interface. The intended audience should be considered in conjunction with the communication protocol to choose a suitable explanation type [36, property O6]. Domain experts may prefer explanations expressed in terms of the internal parameters of the underlying predictive model, but a lay audience may rather prefer exemplar explanations that use relevant data points—choosing the appropriate explanation domain [36, property O4]. The audience also determines the purpose of the explanation. For example, inspecting a predictive model for debugging purposes will need a different system than guiding the explainee with an actionable advice towards a certain goal like getting a loan. Interactive explainers can support a wide spectrum of these properties by allowing the explainee to personalise the output of the explainer as discussed in Sect. 3.2.
Achieving some of these objectives may require the features of the underlying data set or the predictive model itself to be transparent [36, property O5]. For example, consider explaining a model trained on a data set with features that are object measurements in meters in contrast to magnitudes of embedding vectors. When the raw features (original domain) are not human-interpretable, the system designer may decide to use an interpretable representation (transformed domain) to aid the explainee. Providing the users with the provenance of an explanation may help them to better understanding its origin, e.g., an explanation purely based on data, model parameters or both [36, property O10]. Choosing the right explanation family is also important, for example: relation between data features and the prediction, relevant examples such as similar data points or causal mechanisms [36, property O1]. Again, interactive explainers have the advantage of giving the user the opportunity to switch between multiple different explanation types. Furthermore, the design of the user interface should be grounded in the Interactive Machine Learning, Human–Computer Interaction, User Experience and Explainable Artificial Intelligence research to seemingly deliver the explanations. For example, the explainee should be given the opportunity to reverse the effect of any actions that may influence the internal state of the explainer and the system should always respect user’s preferences and feedback [36, property U4]. Finally, if an explanation of the same event can change over time or is influenced by a random factor, user’s trust is at stake. The explainee should always be informed about the degree of explanation invariance and its manifestation in the output of an explainer [36, property S3]. This property is vital to Glass-Box’s success, which we discuss in more detail in Sect. 4.
Glass-Box Reception and Feedback
We presented Glass-Box to domain experts (general AI background knowledge) and a lay audience with the intention to gauge their reception of our prototype and collect feedback that would help us revise and improve our explainability system. To this end, we opted for informal and unstructured free-form feedback, which was mostly user-driven and guided by reference questions (based on our list of desiderata) whenever necessary. We decided to take this approach given the nature of the events at which we presented our prototype—a scientific conference and a research festival.
Glass-Box is composed of multiple independent components, all of which play a role in the user’s reception of the system:
natural language understanding and generation,
speech transcription and synthesis,
voice and text user interfaces, and
a data set that determines the problem domain.
Therefore, collecting free-form feedback at this early stage helped us to pin-point components of the system that required more attention and identify possible avenues for formal testing and design of user studies.
While presenting the device we only approached members of the audience who expressed an interest in interacting with the device and who afterwards were willing to describe their experience. In total, we collected feedback from 6 domain experts and 11 participants of the research festival of varying demographics. When introducing the system and its modes of operation to the participants, we assessed their level of AI and ML expertise by asking background questions, which allowed us to appropriately structure the feedback session.
While discussing the system with the participants, we were mainly interested in their perception of its individual components and suggestions about how these can be improved. Most of the participants enjoyed asking questions and interacting with the device via the voice interface, however some of them found the speech synthesis module that answered their questions “slow”, “unnatural” and “clunky”. These observations have prompted some of the participants to disable voice-based responses and use the text-based chat interface to read the answers instead of listening to them. When asked about the quality of explanations, their comprehensibility and content, many participants were satisfied with received answers. They claimed that personalised explanations provided them with information that they were seeking for as opposed to the default explanation given at first. However, some of them expressed concerns regarding the deployment of such systems in everyday life and taking the human out of the loop. The most common worry was the impossibility to “argue” and “convince” the explainer that the decision is incorrect and the explanation does not capture the complexity of one’s case. Some participants were also sceptical of the general idea of interacting with an AI agent and the fail-safe mode of the device, which produced “I cannot help you with this query.” response whenever the explainer could not answer the user’s question.
We plan to use all of this feedback and our experience in building interactive explainers to refine the system focusing on its explanation personalisation aspect, and test this particular component with formal user studies. Isolating this module of the explainer will alleviate the influence of the user interface on the perception of the explanations, allowing us to investigate the effectiveness and reception of personalised explanations in a formal setting.