Research Papers

Explanations in XAI are typically developed by AI experts and focus on algorithmic transparency and the inner workings of AI systems. Research has shown that such explanations do not meet the needs of users who do not have AI expertise. As a result, explanations are often ineffective in making system decisions interpretable and understandable. We aim to strengthen a socio-technical view of AI by following a Human-Centered Explainable Artificial Intelligence (HC-XAI) approach, which investigates the explanation needs of end-users (i.e., subject matter experts and lay users) in specific usage contexts. One of the most influential works in this area is the XAI Question Bank (XAIQB) by Liao et al. The authors propose a set of questions that end-users might ask when using an AI system, which in turn is intended to help developers and designers identify and address explanation needs. Although the XAIQB is widely referenced, there are few reports of its use in practice. In particular, it is unclear to what extent the XAIQB sufficiently captures the explanation needs of end-users and what potential problems exist in the practical application of the XAIQB. To explore these open questions, we used the XAIQB as the basis for analyzing 12 think-aloud software explorations with subject matter experts. We investigated the suitability of the XAIQB as a tool for identifying explanation needs in a specific usage context. Our analysis revealed a number of explanation needs that were missing from the question bank, but that emerged repeatedly as our study participants interacted with an AI system. We also found that some of the XAIQB questions were difficult to distinguish and required interpretation during use. Our contribution is an extension of the XAIQB with 11 new questions. In addition, we have expanded the descriptions of all new and existing questions to facilitate their use.

In recent years, the community of 'explainable artificial intelligence' (XAI) has created a vast body of methods to bridge a perceived gap between model 'complexity' and 'interpretability'. However, a concrete problem to be solved by XAI methods has not yet been formally stated. As a result, XAI methods are lacking theoretical and empirical evidence for the 'correctness' of their explanations, limiting their potential use for quality-control and transparency purposes. At the same time, Haufe et al. (2014) showed, using simple toy examples, that even standard interpretations of linear models can be highly misleading. Specifically, high importance may be attributed to so-called suppressor variables lacking any statistical relation to the prediction target. This behavior has been confirmed empirically for a large array of XAI methods in Wilming et al. (2022). Here, we go one step further by deriving analytical expressions for the behavior of a variety of popular XAI methods on a simple two-dimensional binary classification problem involving Gaussian class-conditional distributions. We show that the majority of the studied approaches will attribute non-zero importance to a non-class-related suppressor feature in the presence of correlated noise. This poses important limitations on the interpretations and conclusions that the outputs of these XAI methods can afford.

Explainable AI (XAI) techniques have been widely used to help explain and understand the output of deep learning models in fields such as image classification and Natural Language Processing. Interest in using XAI techniques to explain deep learning-based automatic speech recognition (ASR) is emerging. but there is not enough evidence on whether these explanations can be trusted. To address this, we adapt a state-of-the-art XAI technique from the image classification domain, Local Interpretable Model-Agnostic Explanations (LIME), to a model trained for a TIMIT-based phoneme recognition task. This simple task provides a controlled setting for evaluation while also providing expert annotated ground truth to assess the quality of explanations. We find a variant of LIME based on time partitioned audio segments, that we propose in this paper, produces the most reliable explanations, containing the ground truth 96% of the time in its top three audio segments.

For the purpose of explaining multivariate outlyingness, it is shown that the squared Mahalanobis distance of an observation can be decomposed into outlyingness contributions originating from single variables. The decomposition is obtained using the Shapley value, a well-known concept from game theory that became popular in the context of Explainable AI. In addition to outlier explanation, this concept also relates to the recent formulation of cellwise outlyingness, where Shapley values can be employed to obtain variable contributions for outlying observations with respect to their "expected" position given the multivariate data structure. In combination with squared Mahalanobis distances, Shapley values can be calculated at a low numerical cost, making them even more attractive for outlier interpretation. Simulations and real-world data examples demonstrate the usefulness of these concepts.

Many researchers and policymakers have expressed excitement about algorithmic explanations enabling more fair and responsible decision-making. However, recent experimental studies have found that explanations do not always improve human use of algorithmic advice. In this study, we shed light on how people interpret and respond to counterfactual explanations (CFEs) -- explanations that show how a model's output would change with marginal changes to its input(s) -- in the context of pretrial risk assessment instruments (PRAIs). We ran think-aloud trials with eight sitting U.S. state court judges, providing them with recommendations from a PRAI that includes CFEs. We found that the CFEs did not alter the judges' decisions. At first, judges misinterpreted the counterfactuals as real -- rather than hypothetical -- changes to defendants. Once judges understood what the counterfactuals meant, they ignored them, stating their role is only to make decisions regarding the actual defendant in question. The judges also expressed a mix of reasons for ignoring or following the advice of the PRAI without CFEs. These results add to the literature detailing the unexpected ways in which people respond to algorithms and explanations. They also highlight new challenges associated with improving human-algorithm collaborations through explanations.

Counterfactual Evaluation for Explainable AI

Yingqiang Ge, Shuchang Liu, Zelong Li +6

While recent years have witnessed the emergence of various explainable methods in machine learning, to what degree the explanations really represent the reasoning process behind the model prediction -- namely, the faithfulness of explanation -- is still an open problem. One commonly used way to measure faithfulness is \textit{erasure-based} criteria. Though conceptually simple, erasure-based criterion could inevitably introduce biases and artifacts. We propose a new methodology to evaluate the faithfulness of explanations from the \textit{counterfactual reasoning} perspective: the model should produce substantially different outputs for the original input and its corresponding counterfactual edited on a faithful feature. Specially, we introduce two algorithms to find the proper counterfactuals in both discrete and continuous scenarios and then use the acquired counterfactuals to measure faithfulness. Empirical results on several datasets show that compared with existing metrics, our proposed counterfactual evaluation method can achieve top correlation with the ground truth under diffe

Algorithmic Bias and Data Bias: Understanding the Relation between Distributionally Robust Optimization and Data Curation

Agnieszka Słowik, Léon Bottou

Machine learning systems based on minimizing average error have been shown to perform inconsistently across notable subsets of the data, which is not exposed by a low average error for the entire dataset. In consequential social and economic applications, where data represent people, this can lead to discrimination of underrepresented gender and ethnic groups. Given the importance of bias mitigation in machine learning, the topic leads to contentious debates on how to ensure fairness in practice (data bias versus algorithmic bias). Distributionally Robust Optimization (DRO) seemingly addresses this problem by minimizing the worst expected risk across subpopulations. We establish theoretical results that clarify the relation between DRO and the optimization of the same loss averaged on an adequately weighted training dataset. The results cover finite and infinite number of training distributions, as well as convex and non-convex loss functions. We show that neither DRO nor curating the training set should be construed as a complete solution for bias mitigation: in the same way that there is no universally robust training set, there is no universal way to setup a DRO problem and ensure a socially acceptable set of results. We then leverage these insights to provide a mininal set of practical recommendations for addressing bias with DRO. Finally, we discuss ramifications of our results in other related applications of DRO, using an example of adversarial robustness. Our results show that there is merit to both the algorithm-focused and the data-focused side of the bias debate, as long as arguments in favor of these positions are precisely qualified and backed by relevant mathematics known today.

While deep learning (DL) approaches are reaching human-level performance for many tasks, including for diagnostics AI, the focus is now on challenges possibly affecting DL deployment, including AI privacy, domain generalization, and fairness. This last challenge is addressed in this study. Here we look at a novel method for ensuring AI fairness with respect to protected or sensitive factors. This method uses domain adaptation via training set enhancement to tackle bias-causing training data imbalance. More specifically, it uses generative models that allow the generation of more synthetic training samples for underrepresented populations. This paper applies this method to the use case of detection of age related macular degeneration (AMD). Our experiments show that starting with an originally biased AMD diagnostics model the method has the ability to improve fairness.

We propose a novel method for enforcing AI fairness with respect to protected or sensitive factors. This method uses a dual strategy performing training and representation alteration (TARA) for the mitigation of prominent causes of AI bias by including: a) the use of representation learning alteration via adversarial independence to suppress the bias-inducing dependence of the data representation from protected factors; and b) training set alteration via intelligent augmentation to address bias-causing data imbalance, by using generative models that allow the fine control of sensitive factors related to underrepresented populations via domain adaptation and latent space manipulation. When testing our methods on image analytics, experiments demonstrate that TARA significantly or fully debiases baseline models while outperforming competing debiasing methods that have the same amount of information, e.g., with (% overall accuracy, % accuracy gap) = (78.8, 0.5) vs. the baseline method's score of (71.8, 10.5) for EyePACS, and (73.7, 11.8) vs. (69.1, 21.7) for CelebA. Furthermore, recognizing certain limitations in current metrics used for assessing debiasing performance, we propose novel conjunctive debiasing metrics. Our experiments also demonstrate the ability of these novel metrics in assessing the Pareto efficiency of the proposed methods.

Certainty around the regulatory environment is crucial to enable responsible AI innovation and foster the social acceptance of these powerful new technologies. One notable source of uncertainty is, however, that the existing legal liability system is inapt to assign responsibility where a potentially harmful conduct and/or the harm itself are unforeseeable, yet some instantiations of AI and/or the harms they may trigger are not foreseeable in the legal sense. The unpredictability of how courts would handle such cases makes the risks involved in the investment and use of AI incalculable, creating an environment that is not conducive to innovation and may deprive society of some of the benefits AI could provide. To tackle this problem, we propose to draw insights from financial regulatory best-practices and establish a system of AI guarantee schemes. We envisage the system to form part of the broader market-structuring regulatory framework, with the primary function to provide a readily available, clear, and transparent funding mechanism to compensate claims that are either extremely hard or impossible to realize via conventional litigation. We propose it to be at least partially industry-funded, with funding arrangements depending on whether it would pursue other potential policy goals. We aim to engage in a high-level, comparative conceptual debate around the suitability of the foreseeability concept to limit legal liability rather than confronting the intricacies of the case law of specific jurisdictions. Recognizing the importance of the latter task, we leave this to further research in support of the legal system's incremental adaptation to the novel challenges of present and future AI technologies.

Artificial Intelligence Fairness in the Context of Accessibility Research on Intelligent Systems for People who are Deaf or Hard of Hearing

Sushant Kafle, Abraham Glasser, Sedeeq Al-khazraji +3

We discuss issues of Artificial Intelligence (AI) fairness for people with disabilities, with examples drawn from our research on human-computer interaction (HCI) for AI-based systems for people who are Deaf or Hard of Hearing (DHH). In particular, we discuss the need for inclusion of data from people with disabilities in training sets, the lack of interpretability of AI systems, ethical responsibilities of access technology researchers and companies, the need for appropriate evaluation metrics for AI-based access technologies (to determine if they are ready to be deployed and if they can be trusted by users), and the ways in which AI systems influence human behavior and influence the set of abilities needed by users to successfully interact with computing systems.

TIMING: Temporality-Aware Integrated Gradients for Time Series Explanation

Hyeongwon Jang, Changhun Kim, Eunho Yang

Recent explainable artificial intelligence (XAI) methods for time series primarily estimate point-wise attribution magnitudes, while overlooking the directional impact on predictions, leading to suboptimal identification of significant points. Our analysis shows that conventional Integrated Gradients (IG) effectively capture critical points with both positive and negative impacts on predictions. However, current evaluation metrics fail to assess this capability, as they inadvertently cancel out opposing feature contributions. To address this limitation, we propose novel evaluation metrics-Cumulative Prediction Difference (CPD) and Cumulative Prediction Preservation (CPP)-to systematically assess whether attribution methods accurately identify significant positive and negative points in time series XAI. Under these metrics, conventional IG outperforms recent counterparts. However, directly applying IG to time series data may lead to suboptimal outcomes, as generated paths ignore temporal relationships and introduce out-of-distribution samples. To overcome these challenges, we introduce TIMING, which enhances IG by incorporating temporal awareness while maintaining its theoretical properties. Extensive experiments on synthetic and real-world time series benchmarks demonstrate that TIMING outperforms existing time series XAI baselines. Our code is available at https://github.com/drumpt/TIMING.