Abstract Rectified linear unit (ReLU) based neural networks (NNs) are recognised for their remarkable accuracy. However, the decision-making processes of these networks are often complex and difficult to understand. This complexity can lead to challenges in error identification, establishing trust, and conducting thorough analyses. Existing methods often fail to provide clear insights into the actual computations occurring within each layer of these networks. To address this challenge, this study introduces a mechanistic interpretability method called ReLU Region Reasoning (Re3). This method uses the known piecewise-linear characteristics of ReLU networks to offer insights into neuron activation and accurately assess how each feature contributes to the final output and probability. Re3 effectively determines neuron activations and evaluates the contribution of each feature within a specified linear region. Experiments conducted on multiple benchmark datasets, including both tabular and image data, demonstrate that Re3 can replicate individual predictions without error, align feature importance with domain expertise, and maintain consistency with current explanatory methods, thereby avoiding the typical randomness. Analysing neurons reveals activation sparsity and identifies dominant units, thus providing clear targets for model simplification and troubleshooting. By ensuring transparency and algebraic accessibility in each stage of a ReLU-based NN’s decision process, Re3 can be a valuable practical tool for achieving precise mechanistic interpretability.

• A privacy-preserving synthetic ground-truth data generation method is proposed for additive feature attribution evaluation. • Additive Case-based Reasoning (AddCBR) is introduced as a model-aligned baseline for evaluating additive feature attribution. • Coefficient of Quartile Variation (CQV) is applied for the first time to assess the quality of feature attribution. • A complete evaluation pipeline for additive feature attribution is developed to benchmark regression explainability. • The novelty of the work is demonstrated for a real-world regression problem of flight take-off delay prediction. Explainable artificial intelligence (XAI) methods produce information outputs based on a target artificial intelligence model to be explained. The most popular information output is produced by XAI methods of the category feature attribution, which produce the relative contribution of each input feature in a local instance. These relative contributions indicate how important each input feature is in a decision; this type of information is expected to provide explanatory value to users. In real-world regression tasks, feature attribution methods are crucial for comprehending model predictions. However, robust evaluation of such methods remains challenging due to a lack of ground-truth data and widely accepted evaluation metrics, such as accuracy for classification or mean absolute error for regression. This paper proposes a novel approach for generating synthetic, privacy-preserving ground-truth datasets for regression problems that retain original feature behaviour, enabling rigorous feature attribution evaluation without compromising sensitive information. We introduce additive case-based reasoning (AddCBR) as a model-aligned and interpretable baseline to benchmark additive feature attribution methods. This work also demonstrates the first use of the coefficient of quartile variation (CQV) as a statistical measure to quantify the consistency and stability of feature attribution methods. Altogether, these contributions form a comprehensive evaluation methodology for objectively assessing and comparing feature attribution methods in regression models. By providing a controlled evaluation pipeline with reliable baselines and metrics, this work addresses the current lack of consensus and benchmarking in XAI evaluation for regression models.

Artificial Intelligence (AI) has recently made significant advancements and is now pervasive across various application domains. This holds true for Air Transportation as well, where AI is increasingly involved in decision-making processes. While these algorithms are designed to assist users in their daily tasks, they still face challenges related to acceptance and trustworthiness. Users often harbor doubts about the decisions proposed by AI, and in some cases, they may even oppose them. This is primarily because AI-generated decisions are often opaque, non-intuitive, and incompatible with human reasoning. Moreover, when AI is deployed in safety-critical contexts like Air Traffic Management (ATM), the individual decisions generated by AI models must be highly reliable for human operators. Understanding the behavior of the model and providing explanations for its results are essential requirements in every life-critical domain. In this scope, this project aimed to enhance transparency and explainability in AI algorithms within the Air Traffic Management domain. This article presents the results of the project’s validation conducted for a Conflict Detection and Resolution task involving 21 air traffic controllers (10 experts and 11 students) in En-Route position (i.e. hight altitude flight management). Through a controlled study incorporating three levels of explanation, we offer initial insights into the impact of providing additional explanations alongside a conflict resolution algorithm to improve decision-making. At a high level, our findings indicate that providing explanations is not always necessary, and our project sheds light on potential research directions for education and training purposes.

Tree-ensemble models, such as Extreme Gradient Boosting (XGBoost), are renowned Machine Learning models which have higher prediction accuracy compared to traditional tree-based models. This higher accuracy, however, comes at the cost of reduced interpretability. Also, the decision path or prediction rule of XGBoost is not explicit like the tree-based models. This paper proposes the iXGB--interpretable XGBoost, an approach to improve the interpretability of XGBoost. iXGB approximates a set of rules from the internal structure of XGBoost and the characteristics of the data. In addition, iXGB generates a set of counterfactuals from the neighbourhood of the test instances to support the understanding of the end-users on their operational relevance. The performance of iXGB in generating rule sets is evaluated with experiments on real and benchmark datasets which demonstrated reasonable interpretability. The evaluation result also supports that the interpretability of XGBoost can be improved without using surrogate methods.

Automated analyses of electroencephalographic (EEG) signals acquired in naturalistic environments are becoming increasingly important in areas such as brain-computer interfaces and behavior science. However, the recorded EEG in such environments is often heavily contaminated by motion artifacts and eye movements. This poses new requirements on artifact handling. The objective of this paper is to present an automated EEG artifacts handling algorithm, which will be used as a preprocessing step in a driver monitoring application. The algorithm, named Automated aRTifacts handling in EEG (ARTE), is based on wavelets, independent component analysis, and hierarchical clustering. The algorithm is tested on a dataset obtained from a driver sleepiness study including 30 drivers and 540 30-min 30-channel EEG recordings. The algorithm is evaluated by a clinical neurophysiologist, by quantitative criteria (signal quality index, mean square error, relative error, and mean absolute error), and by demonstrating its usefulness as a preprocessing step in driver monitoring, here exemplified with driver sleepiness classification. All results are compared with a state-of-the-art algorithm called FORCe. The quantitative and expert evaluation results show that the two algorithms are comparable, and that both algorithms significantly reduce the impact of artifacts in recorded EEG signals. When artifact handling is used as a preprocessing step in driver sleepiness classification, the classification accuracy increased by 5% when using ARTE and by 2% when using FORCe. The advantage with ARTE is that it is data driven and does not rely on additional reference signals or manually defined thresholds, making it well suited for use in dynamic settings where unforeseen and rare artifacts are commonly encountered.

The health sciences are, nowadays, one of the major application areas for case-based reasoning (CBR). The paper presents a survey of recent medical CBR systems based on a literature review and an e-mail questionnaire sent to the corresponding authors of the papers where these systems are presented. Some clear trends have been identified, such as multipurpose systems: more than half of the current medical CBR systems address more than one task. Research on CBR in the area is growing, but most of the systems are still prototypes and not available in the market as commercial products. However, many of the projects/systems are intended to be commercialized.

Stress diagnosis based on finger temperature (FT) signals is receiving increasing interest in the psycho‐physiological domain. However, in practice, it is difficult and tedious for a clinician and particularly less experienced clinicians to understand, interpret, and analyze complex, lengthy sequential measurements to make a diagnosis and treatment plan. The paper presents a case‐based decision support system to assist clinicians in performing such tasks. Case‐based reasoning (CBR) is applied as the main methodology to facilitate experience reuse and decision explanation by retrieving previous similar temperature profiles. Further fuzzy techniques are also employed and incorporated into the CBR system to handle vagueness, uncertainty inherently existing in clinicians reasoning as well as imprecision of feature values. Thirty‐nine time series from 24 patients have been used to evaluate the approach (matching algorithms) and an expert has ranked and estimated similarity. On average goodness‐of‐fit for the fuzzy matching algorithm is 90% in ranking and 81% in similarity estimation that shows a level of performance close to an experienced expert. Therefore, we have suggested that a fuzzy matching algorithm in combination with CBR is a valuable approach in domains, where the fuzzy matching model similarity and case preference is consistent with the views of domain expert. This combination is also valuable, where domain experts are aware that the crisp values they use have a possibility distribution that can be estimated by the expert and is used when experienced experts reason about similarity. This is the case in the psycho‐physiological domain and experienced experts can estimate this distribution of feature values and use them in their reasoning and explanation process.