Artificial intelligence (AI) systems are increasingly adopted in high‐stakes domains such as healthcare and finance, so the demand for transparency and interpretability has grown substantially. EXplainable AI (XAI) methods have emerged to address this challenge, but individual techniques often offer limited, fragmented insights. This paper introduces Meta‐explainers, a novel ensemble‐based XAI framework that integrates multiple explanation types—specifically relevance‐based and counterfactual methods—into unified, multifaceted and complementary meta‐explanations. Inspired by meta‐classification principles, our approach structures the explanation process into five stages: generation, grouping, evaluation, aggregation, and visualization. Each stage is designed to preserve the unique strengths of individual XAI techniques while enhancing their interpretability and coherence when combined. Experimental results on both image (MNIST) and tabular (Breast Cancer) datasets show that Meta‐explainers consistently outperform individual and state‐of‐the‐art ensemble explanation methods in terms of explanation quality, as measured by established metrics. This work paves the way toward more holistic and user‐centered AI explainability with a flexible methodology that can be extended to incorporate additional explanation paradigms.

Explainability has become one of the leading research topics within Artificial Intelligence (AI) in the last few years, as it has increased the confidence and credibility of “black box” models, such as deep neural networks . However, the evaluation of the explanations provided by different explainability approaches remains a hot research topic. This evaluation enables the possibility of comparing different existing techniques and would play a crucial role in improving the auditability of AI-based systems. The current literature on the subject considers that the evaluation of explainability methods can be approached in two ways: qualitative and quantitative . While qualitative evaluations are based on assumptions induced by human understanding that are hard to develop and prone to introduce certain cognitive biases, quantitative ones avoid these biases by excluding the human expert from the evaluation process. However, the main challenge in quantitatively evaluating an explanation is the lack of ground truth specifying what defines a correct explanation. In this paper, we propose an evaluation measure that quantifies the Level of Strength of an Explanation (LSE), i.e., the extent to which the explanation produced by a post-hoc explainability method supports the class predicted by a classifier. Our proposal is inspired by the semantics underlying the Likelihood Ratio in evaluating forensic evidence, which is defined as the weight to be attributed to a piece of forensic evidence according to the prosecution and defense propositions. To validate our proposal, nine popular explainability techniques are compared across two deep neural architectures dedicated to image classification and three classifiers for binary classification over tabular datasets. In addition, we use MetaQuantus as a meta-evaluation approach. Results from our experimental study reveal that GradCAM and LRP outperform the other explainability methods in terms of the proposed LSE measure.

Although deep learning models can solve complex prediction problems, they have been criticized for being ‘black boxes’. This implies that their decisions are difficult, if not impossible, to explain by simply inspecting their internal knowledge structures. Explainable Artificial Intelligence has attempted to open the black-box through model-specific and agnostic post-hoc methods that generate visualizations or derive associations between the problem features and the model predictions. This paper proposes a new method, termed REPROT, that explains the decisions of complex deep learning architectures based on local reducts of an image. A ‘reduct’ is a set of sufficiently descriptive features that can fully characterize the acquired knowledge. The created reducts are used to build a ‘prototype image’ that visually explains the inference obtained by a black-box model for an image. We focus on deep learning architectures whose complexity and internal particularities demand adapting existing model-specific explanation methods, making the explanation process more difficult. Experimental results show that the black-box model can detect an object using the prototype image generated from the reduct. Hence, the explanations will be given by “the minimum set of features sufficient for the neural model to detect an object”. The confidence scores obtained by architectures such as Inception, Yolo, and Mask R-CNN are higher for prototype images built from the reduct than those built from the most important superpixels according to the LIME method. Moreover, the target object is not detected on several occasions through the LIME output, thus supporting the superiority of the proposed explanation method.

Neural networks are considered a black-box model as their strength in modeling complex interactions makes its operation almost impossible to explain. Still, neural networks remain very interesting tools as they have shown promising performance in various classification tasks. Layer-wise relevance propagation is a technique that, based on a propagation approach, is able to explain the predictions obtained by a neural network. In this work, we propose four adaptations of this technique to operate on multi-label neural networks. The proposed methods provide new ways of distributing the relevance between the output layer and the preceding ones. The efficacy of these adaptations is demonstrated after an experimental study. The study is carried out based on existing evaluation criteria in the literature that measure the explanation's quality. These methods are applied to a case study in which a neural network is used to detect secondary coinfections in patients infected with SARS-CoV-2. Overall, the proposed methods provide a post-hoc interpretability stage of the results.

This paper presents a neural system to deal with multi-label classification problems that might involve sparse features. The architecture of this model involves three sequential blocks with well-defined functions. The first block consists of a multilayered feed-forward structure that extracts hidden features, thus reducing the problem dimensionality. This block is useful when dealing with sparse problems. The second block consists of a Long-term Cognitive Network-based model that operates on features extracted by the first block. The activation rule of this recurrent neural network is modified to prevent the vanishing of the input signal during the recurrent inference process. The modified activation rule combines the neurons' state in the previous abstract layer (iteration) with the initial state. Moreover, we add a bias component to shift the transfer functions as needed to obtain good approximations. Finally, the third block consists of an output layer that adapts the second block's outputs to the label space. We propose a backpropagation learning algorithm that uses a squared hinge loss function to maximize the margins between labels to train this network. The results show that our model outperforms the state-of-the-art algorithms in most datasets.