How much can anyone trust a recommendation from an AI? Yaroslav Kuflinski, from Iflexion gives an explanation of explainable AI She is lying sedated on a gurney that's bumping towards the operating theater. It squeaks to a halt and a hurried member of hospital staff thrusts a form at you to sign. It describes the urgent surgical procedure your child is about to undergo--and it requires your signature if the operation is to go ahead. At this specific moment, do you think you are owed a reasonable, plain-English explanation of all the inscrutable decisions that an AI has lately been making on your daughter's behalf? in short, do we need explainable AI? There are many other examples where one or more of the actors may consider themselves entitled to an explanation of the reasoning processes behind the decisions of an AI.

We predict credit applications with off-the-shelf, interchangeable black-box classifiers and we explain single predictions with counterfactual explanations. Counterfactual explanations expose the minimal changes required on the input data to obtain a different result e.g., approved vs rejected application. Despite their effectiveness, counterfactuals are mainly designed for changing an undesired outcome of a prediction i.e. loan rejected. Counterfactuals, however, can be difficult to interpret, especially when a high number of features are involved in the explanation. Our contribution is two-fold: i) we propose positive counterfactuals, i.e. we adapt counterfactual explanations to also explain accepted loan applications, and ii) we propose two weighting strategies to generate more interpretable counterfactuals. Experiments on the HELOC loan applications dataset show that our contribution outperforms the baseline counterfactual generation strategy, by leading to smaller and hence more interpretable counterfactuals.

Mathematical optimization offers highly-effective tools for finding solutions for problems with well-defined goals, notably scheduling. However, optimization solvers are often unexplainable black boxes whose solutions are inaccessible to users and which users cannot interact with. We define a novel paradigm using argumentation to empower the interaction between optimization solvers and users, supported by tractable explanations which certify or refute solutions. A solution can be from a solver or of interest to a user (in the context of 'what-if' scenarios). Specifically, we define argumentative and natural language explanations for why a schedule is (not) feasible, (not) efficient or (not) satisfying fixed user decisions, based on models of the fundamental makespan scheduling problem in terms of abstract argumentation frameworks (AFs). We define three types of AFs, whose stable extensions are in one-to-one correspondence with schedules that are feasible, efficient and satisfying fixed decisions, respectively. We extract the argumentative explanations from these AFs and the natural language explanations from the argumentative ones.

The paper develops a formal theory of the degree of justification of arguments, which relies solely on the structure of an argumentation framework, and which can be successfully interfaced with approaches to instantiated argumentation. The theory is developed in three steps. First, the paper introduces a graded generalization of the two key notions underpinning Dung's semantics: self-defense and conflict-freeness. This leads to a natural generalization of Dung's semantics, whereby standard extensions are weakened or strengthened depending on the level of self-defense and conflict-freeness they meet. The paper investigates the fixpoint theory of these semantics, establishing existence results for them. Second, the paper shows how graded semantics readily provide an approach to argument rankings, offering a novel contribution to the recently growing research programme on ranking-based semantics. Third, this novel approach to argument ranking is applied and studied in the context of instantiated argumentation frameworks, and in so doing is shown to account for a simple form of accrual of arguments within the Dung paradigm. Finally, the theory is compared in detail with existing approaches.

The topic of causal explanation in artificial intelligence has gathered interest in recent years as researchers and practitioners aim to increase trust and understanding of intelligent decision-making and action. While different sub-fields have looked into this problem with a sub-field-specific view, there are few models that aim to capture explanation in AI more generally. One general model is based on structural causal models. It defines an explanation as a fact that, if found to be true, would constitute an actual cause of a specific event. However, research in philosophy and social sciences shows that explanations are contrastive: that is, when people ask for an explanation of an event -- the fact --- they (sometimes implicitly) are asking for an explanation relative to some contrast case; that is, "Why P rather than Q?". In this paper, we extend the structural causal model approach to define two complementary notions of contrastive explanation, and demonstrate them on two classical AI problems: classification and planning. We believe that this model can be used to define contrastive explanation of other subfield-specific AI models.

The act of persuasion, a key component in rhetoric argumentation, may be viewed as a dynamics modifier. We extend Dung's frameworks with acts of persuasion among agents, and consider interactions among attack, persuasion and defence that have been largely unheeded so far. We characterise basic notions of admissibilities in this framework, and show a way of enriching them through, effectively, CTL (computation tree logic) encoding, which also permits importation of the theoretical results known to the logic into our argumentation frameworks. Our aim is to complement the growing interest in coordination of static and dynamic argumentation.

Recent work on interpretability in machine learning and AI has focused on the building of simplified models that approximate the true criteria used to make decisions. These models are a useful pedagogical device for teaching trained professionals how to predict what decisions will be made by the complex system, and most importantly how the system might break. However, when considering any such model it's important to remember Box's maxim that "All models are wrong but some are useful." We focus on the distinction between these models and explanations in philosophy and sociology. These models can be understood as a "do it yourself kit" for explanations, allowing a practitioner to directly answer "what if questions" or generate contrastive explanations without external assistance. Although a valuable ability, giving these models as explanations appears more difficult than necessary, and other forms of explanation may not have the same trade-offs. We contrast the different schools of thought on what makes an explanation, and suggest that machine learning might benefit from viewing the problem more broadly.

Building explainable systems is a critical problem in the field of Natural Language Processing (NLP), since most machine learning models provide no explanations for the predictions. Existing approaches for explainable machine learning systems tend to focus on interpreting the outputs or the connections between inputs and outputs. However, the fine-grained information is often ignored, and the systems do not explicitly generate the human-readable explanations. To better alleviate this problem, we propose a novel generative explanation framework that learns to make classification decisions and generate fine-grained explanations at the same time. More specifically, we introduce the explainable factor and the minimum risk training approach that learn to generate more reasonable explanations. We construct two new datasets that contain summaries, rating scores, and fine-grained reasons. We conduct experiments on both datasets, comparing with several strong neural network baseline systems. Experimental results show that our method surpasses all baselines on both datasets, and is able to generate concise explanations at the same time.

Artificial Intelligence (AI) systems are taking over a vast array of tasks that previously depended on human expertise and judgment. Often, however, the "reasoning" behind their actions is unclear, and can produce surprising errors or reinforce biased processes. One way to address this issue is to make AI "explainable" to humans--for example, designers who can improve it or let users better know when to trust it. Although the best styles of explanation for different purposes are still being studied, they will profoundly shape how future AI is used. Some explainable AI, or XAI, has long been familiar, as part of online recommender systems: book purchasers or movie viewers see suggestions for additional selections described as having certain similar attributes, or being chosen by similar users.

This paper introduces the "grasp-ability test" as a "goodness" criteria by which to compare which explanation is more or less meaningful than others for users to understand the automated algorithmic data processing. A growing number of researchers attempt to develop explainable AIs (hereafter, XAI) to meet practical (e.g., explainability is positively correlated to users' learning performance), legal (e.g., explainability is required for S.E.C. to scrutinize AIpowered trading techniques; liability issues) and ethical expectations (e.g., right to explanation; trust; autonomy). Different researchers use different ideas of what an explanation is [1]. For example, as Figure 1 shows, 11 U.S. research groups, funded by DARPA, are currently developing XAI in different manners. Then, a question is raised: how can we know which model of XAI is good enough or better/worse than others? To answer this, we need a "goodness" criteria.