Explainable AI (XAI) methods have mostly been built to investigate and shed light on single machine learning models and are not designed to capture and explain differences between multiple models effectively. This paper addresses the challenge of understanding and explaining differences between machine learning models, which is crucial for model selection, monitoring and lifecycle management in real-world applications. We propose DeltaXplainer, a model-agnostic method for generating rule-based explanations describing the differences between two binary classifiers. To assess the effectiveness of DeltaXplainer, we conduct experiments on synthetic and real-world datasets, covering various model comparison scenarios involving different types of concept drift.

Fuzzy systems are valued for their ability to model complex phenomena while contributing to maintaining inter-pretability through natural language rules. However, this property is not inherently guaranteed by the adoption of fuzzy sets. This paper proposes both a tool, named FuzzIC, to assess the interpretability of fuzzy rule bases through the computation of various complementary criteria taken from the literature, and an empirical user study of these criteria. The latter shows that some criteria capture characteristics that significantly enhance interpretability in practical applications, while others appear less influential than anticipated.

Using multiple-choice questions (MCQs) has become a standard for assessing LLM capabilities efficiently. A variety of metrics can be employed for this task. However, previous research has not conducted a thorough assessment of them. At the same time, MCQ evaluation suffers from answer fluctuation: models produce different results given slight changes in prompts. We suggest a metric assessment protocol in which evaluation methodologies are analyzed through their connection with fluctuation rates, as well as original performance. Our results show that there is a strong link between existing metrics and the answer changing, even when computed without any additional prompt variants. Using the protocol, the highest association is demonstrated by a novel metric, worst accuracy.

In the context of abstract argumentation, we present the benefits of considering temporality, i.e. the order in which arguments are enunciated, as well as causality. We propose a formal method to rewrite the concepts of acyclic abstract argumentation frameworks into an action language that allows us to model the evolution of the world and to establish causal relationships between the enunciation of arguments and their consequences on the acceptability of other arguments, whether direct or indirect. An Answer Set Programming implementation is also described. Furthermore, we lay the ground for the generation of explanations using the causal relations as well as two graphical representations to support such explanations.

When tabular data cannot be directly mined, due to their large size or for privacy reasons, their summary may still be available for analysis. Fuzzy linguistic summaries are composed of sentences describing a multivariate data distribution using, possibly personalized, terms taken from a fuzzy vocabulary. This paper introduces an algorithmic solution to the comparison of summaries so as to provide users with a linguistic description of the data changes between datasets to be compared. A first strategy processes exhaustive summaries containing one sentence for each of the subspaces that can be formed using terms from the vocabulary. A second strategy is proposed for condensed summaries, that involve informative sentences only. Experimentation conducted on artificial datasets confirm the relevance of this second strategy in terms of computational cost and data changes that can be tracked.

Case-based prediction (CBP) methods do not learn a model of the target decision function but instead perform an inference process that depends on two similarity measures and a reference case base. This paper proposes a strategy, called EnergyCompress, to learn an effective case base by selecting relevant cases from an initial set. Use of EnergyCompress decreases CBP inference time, through case base compression, and also increases prediction performance, for a wide variety of CBP algorithms. EnergyCompress relies on the proposition of a general formulation of the CBP task in the framework of energy-based models, which leads to a new and valuable characterization of the notion of competence in case-based reasoning, in particular at the source case level. Extensive experimental results on 18 benchmark datasets comparing EnergyCompress to 5 reference algorithms for case base maintenance support the benefit of the proposed strategy.

This paper proposes Complete Textual Concept Bottleneck Model (CT-CBM), a novel TCBM generator building concept labels in a fully unsupervised manner using a small language model, eliminating both the need for predefined human labeled concepts and LLM annotations. CT-CBM iteratively targets and adds important and identifiable concepts in the bottleneck layer to create a complete concept basis. CT-CBM achieves striking results against competitors in terms of concept basis completeness and concept detection accuracy, offering a promising solution to reliably enhance interpretability of NLP classifiers.

Toxicity mitigation consists in rephrasing text in order to remove offensive or harmful meaning. Neural natural language processing (NLP) models have been widely used to target and mitigate textual toxicity. However, existing methods fail to detoxify text while preserving the initial non-toxic meaning at the same time. In this work, we propose to apply eXplainable AI (XAI) methods to both target and mitigate textual toxicity. We propose CF-Detox _tigtec to perform text detoxification by applying local feature importance, counterfactual example generation and counterfactual feature importance methods to a toxicity classifier distinguishing between toxic and non-toxic texts. We carry out text detoxification through counterfactual generation on three datasets and compare our approach to three competitors. Automatic and human evaluations show that recently developed NLP counterfactual generators lead to competitive results in toxicity mitigation. This work is the first to bridge the gap between counterfactual generation and text detoxification and paves the way towards more practical applications of XAI methods.

A vocabulary materialized by linguistic variables is a way to integrate subjective and expert knowledge about data to analyze. When multiple experts are involved in the analysis process, it is crucial to point out disagreements between their respective vocabularies and the potential impact of this disagreement on the output of a data analysis task applied to the processed data. This paper presents disagreement measures that inform experts about discrepancies regarding the topological structure of their partitions, the terminology associated with each modality, and the impact of the disagreement on the discretization of a given dataset. Experiments show that the proposed multi-level disagreement measure exhibits complementary and meaningful discrepancies that may help experts converge toward a consensual vocabulary.

Large Language Models (LLMs) can generate plausible free text self-explanations to justify their answers. However, these natural language explanations may not accurately reflect the model's actual reasoning process, pinpointing a lack of faithfulness. Existing faithfulness evaluation methods rely primarily on behavioral tests or computational block analysis without examining the semantic content of internal neural representations. This paper proposes NeuroFaith, a flexible framework that measures the faithfulness of LLM free text self-explanation by identifying key concepts within explanations and mechanistically testing whether these concepts actually influence the model's predictions. We show the versatility of NeuroFaith across 2-hop reasoning and classification tasks. Additionally, we develop a linear faithfulness probe based on NeuroFaith to detect unfaithful self-explanations from representation space and improve faithfulness through steering. NeuroFaith provides a principled approach to evaluating and enhancing the faithfulness of LLM free text self-explanations, addressing critical needs for trustworthy AI systems.

The Gödel integral is a variant of the Sugeno integral. It is an expressive aggregation operator to compute a global evaluation combining local values taken for a set of considered criteria and a capacity, a set function that represents the importance of these criteria and their interactions. A crucial question is then the elicitation of this capacity so as to match the user preferences. Such a task is common in many fields, such as XAI, where counterfactual explanations can be generated in the form of counterfactual examples by aggregating a pair of objective and subjective criteria. In this paper, we focus on approaches to elicit such a capacity when the user preferences are expressed as an ordering relation between classes of equivalent objects. We first propose a theoretical characterisation of the set of admissible capacities: it allows to determine lower and upper bounds of possible capacities when it is not empty and it provides explanations when it is empty. Then, we introduce the algorithm GICEP (for Gödel Integral for Capacity Elicitation from Preference relations), to compute these bounds and to provide explanations. Experiments conducted on synthetic data show the relevance and efficiency of GICEP.

Incorporating natural language rationales in the prompt and In-Context Learning (ICL) have led to a significant improvement of Large Language Models (LLMs) performance. However, generating high-quality rationales require human-annotation or the use of auxiliary proxy models. In this work, we propose Self-AMPLIFY to automatically generate rationales from post hoc explanation methods applied to Small Language Models (SLMs) to improve their own performance. Self-AMPLIFY is a 3-step method that targets samples, generates rationales and builds a final prompt to leverage ICL. Self-AMPLIFY performance is evaluated on four SLMs and five datasets requiring strong reasoning abilities. Self-AMPLIFY achieves good results against competitors, leading to strong accuracy improvement. Self-AMPLIFY is the first method to apply post hoc explanation methods to autoregressive language models to generate rationales to improve their own performance in a fully automated manner.

Reconstructing the data inner structure and identifying abnormal points are two major tasks in many data analysis processes. A step beyond the decomposition of a data set as inliers and outliers, that then may be interpreted as anomalies, is to distinguish local from global outliers. This paper introduces a unified approach based on a revised version of an isolation forest that allows for both the reconstruction of dense regions of points, the identification of anomalies and the generation of contextual explanations about the abnormality of these points. To make the anomaly detection more informative and reliable, anomalies are compared to the reconstructed partition of the inliers so as to explain why they are considered abnormal and from which local generation mechanism they could originate from. Relying on a common data property, namely the isolation of anomalies from dense groups of regularities, eases the understanding of the data set structure and makes the provided explanations more informative than those provided by two independent mechanisms, one for clustering and one for detecting anomalies. Conducted experimentations show the relevance of the structural knowledge extracted from the proposed isolation forest and the effectiveness and robustness of the approach thanks to the unified isolation-based data model to analyse different facets of the data.

An abstract argumentation framework is a commonly used formalism to provide a static representation of a dialogue. However, the order of enunciation of the arguments in an argumentative dialogue is very important and can affect the outcome of this dialogue. In this paper, we propose a new framework for modelling abstract argumentation graphs, a model that incorporates the order of enunciation of arguments. By taking this order into account, we have the means to deduce a unique outcome for each dialogue, called an extension. We also establish several properties, such as termination and correctness, and discuss two notions of completeness. In particular, we propose a modification of the previous transformation based on a "last enunciated last updated" strategy, which verifies the second form of completeness.

This paper proposes a tentative and original survey of meeting points between Knowledge Representation and Reasoning (KRR) and Machine Learning (ML), two areas which have been developed quite separately in the last four decades. First, some common concerns are identified and discussed such as the types of representation used, the roles of knowledge and data, the lack or the excess of information, or the need for explanations and causal understanding. Then, the survey is organised in seven sections covering most of the territory where KRR and ML meet. We start with a section dealing with prototypical approaches from the literature on learning and reasoning: Inductive Logic Programming, Statistical Relational Learning, and Neurosymbolic AI, where ideas from rule-based reasoning are combined with ML. Then we focus on the use of various forms of background knowledge in learning, ranging from additional regularisation terms in loss functions, to the problem of aligning symbolic and vector space representations, or the use of knowledge graphs for learning. Then, the next section describes how KRR notions may benefit to learning tasks. For instance, constraints can be used as in declarative data mining for influencing the learned patterns; or semantic features are exploited in low-shot learning to compensate for the lack of data; or yet we can take advantage of analogies for learning purposes. Conversely, another section investigates how ML methods may serve KRR goals. For instance, one may learn special kinds of rules such as default rules, fuzzy rules or threshold rules, or special types of information such as constraints, or preferences. The section also covers formal concept analysis and rough sets-based methods. Yet another section reviews various interactions between Automated Reasoning and ML, such as the use of ML methods in SAT solving to make reasoning faster. Then a section deals with works related to model accountability, including explainability and interpretability, fairness and robustness. Finally, a section covers works on handling imperfect or incomplete data, including the problem of learning from uncertain or coarse data, the use of belief functions for regression, a revision-based view of the EM algorithm, the use of possibility theory in statistics, or the learning of imprecise models. This paper thus aims at a better mutual understanding of research in KRR and ML, and how they can cooperate. The paper is completed by an abundant bibliography.

As Machine Learning models achieve unprecedented levels of performance, the XAI domain aims at making these models understandable by presenting end-users with intelligible explanations. Yet, some existing XAI approaches fail to meet expectations: several issues have been reported in the literature, generally pointing out either technical limitations or misinterpretations by users. In this paper, we argue that the resulting harms arise from a complex overlap of multiple failures in XAI, which existing ad-hoc studies fail to capture. This work therefore advocates for a holistic perspective, presenting a systematic investigation of limitations of current XAI methods and their impact on the interpretation of explanations. By distinguishing between system-specific and user-specific failures, we propose a typological framework that helps revealing the nuanced complexities of explanation failures. Leveraging this typology, we discuss some research directions to help practitioners better understand the limitations of XAI systems and enhance the quality of ML explanations.

Understanding why some points in a data set are considered as anomalies cannot be done without taking into account the structure of the regular points. Whereas many machine learning methods are dedicated to the identification of anomalies on one side, or to the identification of the data inner-structure on the other side, a solution is introduced to answers these two tasks using a same data model, a variant of an isolation forest. The initial algorithm to construct an isolation forest is indeed revisited to preserve the data inner structure without affecting the efficiency of the outlier detection. Experiments conducted both on synthetic and real-world data sets show that, in addition to improving the detection of abnormal data points, the proposed variant of isolation forest allows for a reconstruction of the subspaces of high density. Therefore, the former can serve as a basis for a unified approach to detect global and local anomalies, which is a necessary condition to then provide users with informative descriptions of the data.

The flexibility of the Godel integral as aggregation tool comes from its parameter, defined as a capacity. A crucial question is that of the choice of this parameter. This paper considers the case when two criteria are to be aggregated and provides an elicitation protocol as an incremental algorithm called INES, which stands for Interactive Necessary Example Search. It allows to reduce progressively the uncertainty about the capacity, until it is possible to identify a necessary winner: the latter is an alternative that maximises the Godel integral value for all instantiations induced by the current capacity candidates. The proposed INES algorithm is illustrated in an explainable AI (XAI) context: it is used to find the characteristics of the best counterfactual candidate, in terms of objective and subjective evaluations, even if the user requirements about their aggregation is still not completely known.

This paper revisits the notion of case base competence in the light of recent advances in the modeling of analogical reasoning, based on the idea of similarity transfer from a situation space to an outcome space. For that we consider the CoAT indicator, that measures the compatibility between two similarity measures on a case base, and use it to define an intrinsic measure of competence of a case base with respect to a reference set. Initial experimental results show that the proposed competence measure correlates with the performance of the CoAT prediction algorithm. In fact, our preliminary results seem to indicate that, under some initial conditions, our competence based model can fit any classification boundary. We then revisit the notions of case competence and locality, and show that some source cases may degrade the overall case base competence while others may improve it, and that a given source case may have disparate influence on different regions of the case space