International audience

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.

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.

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.

Plural counterfactual examples have been proposed to explain the prediction of a classifier by offering a user several instances of minimal modifications that may be performed to change the prediction. Yet, such explanations may provide too much information, generating potential confusion for the end-users with no specific knowledge, neither on the machine learning, nor on the application domains. In this paper, we investigate the design of explanation user interfaces for plural counterfactual examples offering comparative analysis features to mitigate this potential confusion and improve the intelligibility of such explanations for non-expert users. We propose an implementation of such an enhanced explanation user interface, illustrating it in a financial scenario related to a loan application. We then present the results of a lab user study conducted with 112 participants to evaluate the effectiveness of having plural examples and of offering comparative analysis principles, both on the objective understanding and satisfaction of such explanations. The results demonstrate the effectiveness of the plural condition, both on objective understanding and satisfaction scores, as compared to having a single counterfactual example. Beside the statistical analysis, we perform a thematic analysis of the participants’ responses to the open-response questions, that also shows encouraging results for the comparative analysis features on the objective understanding.

Explaining a decision requires some properties that have been studied and established in cognitive sciences. An important one is the contrastive nature of explanations: an explanation should answer questions such as "why make decision P rather than Q?". This principle has been formalized recently by T. Miller in a logical framework exploiting knowledge represented as structural causal graphs, with variables taking crisp values. However this framework does not allow us to cope easily with imprecise knowledge or data, nor with imprecise formulations of explanations, that could be preferred in some situations. This paper discusses the principles of such fuzzy extensions of this model, exploring the various levels for integrating fuzzy semantics: it discusses successively (i) the input level, for imprecisely described data instances for which an explanation is required, (ii) the level of the structural causal graph itself, to model imprecise knowledge about the functional relations between the variables involved in the model, and (iii) the output level, to express the explanation, e.g. using fuzzy modalities. The combination of these levels is considered as well. Finally, the paper proposes a discussion about the definition of minimality in the fuzzy framework.

This paper proposes a study of a configurable aggregation operator in many-valued logic (MVL), called ABOP, standing for ABating OPerator. It focuses on its capacity of implementing any t-norm: the paper’s main theorem offers a constructive proof of the existence of a configuration for any given MVL t-norm, showing it is a generalisation of t-norm operators. Furthermore, it specifies the corresponding parameter values. The paper then considers several examples of classical fuzzy logic t-norms adapted to MVL, and examines the corresponding parameters, as instantiation of the general representation theorem.

Linguistic summaries are used in this work to provide personalized exploration functionalities on massive relational data. To ensure a fluid exploration of the data, cardinalities of the data properties described in the summaries are estimated from statistics about the data distribution. The proposed workflow also involves a vocabulary inference mechanism from these statistics and a sampling-based approach to consolidate the estimated cardinalities. The paper shows that soft computing techniques are particularly relevant to build concrete and functional business intelligence solutions.

The increasing usage of complex Machine Learning models for decision-making has raised interest in explainable artificial intelligence (XAI). In this work, we focus on the effects of providing accessible and useful explanations to non-expert users. More specifically, we propose generic XAI design principles for contextualizing and allowing the exploration of explanations based on local feature importance. To evaluate the effectiveness of these principles for improving users’ objective understanding and satisfaction, we conduct a controlled user study with 80 participants using 4 different versions of our XAI system, in the context of an insurance scenario. Our results show that the contextualization principles we propose significantly improve user’s satisfaction and is close to have a significant impact on user’s objective understanding. They also show that the exploration principles we propose improve user’s satisfaction. On the other hand, the interaction of these principles does not appear to bring improvement on both dimensions of users’ understanding.

Subspace clustering can offer, beside a decomposition of data into homogeneous and distinct clusters, a characterisation of the subspaces in which the clusters live. This paper explores the possibility of capturing the notion of characteristic features in the framework of typicality degrees, as typical features. To that aim, it discusses the notion of typicality degrees for features and proposes an Alternating Cluster Estimation algorithm, named TbSC, to exploit these degrees within subspace clustering. It illustrates their differences experimentally using simple data sets.

In the spirit of explainable AI approaches, this paper introduces a new strategy whose aim is to linguistically describe the inner structure of a dataset. Instead of removing irregular points and focusing on the analysis of regular points, the proposed approach relies on a unified data structure, an isolation forest, to both separate regular from irregular points and to identify their inner structure using a data-driven similarity measure. In addition, clusters of regular and irregular points are then linguistically described so as to help users focus on the most characteristic properties of each cluster and to possibly understand the reason why some points are irregular.

The linguistic summarization of a dataset is a process whose complexity depends linearly on the size of the dataset and exponentially on the size of the fuzzy vocabulary. To efficiently summarize large datasets stored in Relational DataBases, reliable estimated cardinalities can be derived from statistics about the data distribution maintained by the RDB Management System, with no expensive data scans. This paper proposes to improve the precision of such estimated summaries while preserving their computational efficiency, by enriching the statistics-based approach with local scan-based corrections when needed: the proposed FRELS method provides efficient strategies both for identifying the needs and performing the corrections. Experiments conducted on real data show that FRELS remains incomparably more efficient than data-scan-based approaches to data summarization and offers a better precision than purely statistics-based approaches. The generation of estimated linguistic summaries takes a couple of seconds, even for datasets containing millions of tuples, with a reliability of more than 95%.

This paper considers a parametrised aggregation operator, originally introduced in the formal framework of many- valued logic and in the applicative context of information scoring. It studies this operator, outside this applicative context, looking at specific configurations of interest: highlighting the wide range of its instantiations, from the lower to the upper extreme cases; showing some t-norms it can encode, as specific cases; and also how it allows rich and flexible intermediate behaviours.

Post-hoc interpretability approaches have been proven to be powerful tools to generate explanations for the predictions made by a trained black-box model. However, they create the risk of having explanations that are a result of some artifacts learned by the model instead of actual knowledge from the data. This paper focuses on the case of counterfactual explanations and asks whether the generated instances can be justified, i.e. continuously connected to some ground-truth data. We evaluate the risk of generating unjustified counterfactual examples by investigating the local neighborhoods of instances whose predictions are to be explained and show that this risk is quite high for several datasets. Furthermore, we show that most state of the art approaches do not differentiate justified from unjustified counterfactual examples, leading to less useful explanations.

Summarizing data with linguistic statements is a crucial and topical issue that has been largely addressed by the soft computing community. The goal of summarization is to generate statements that linguistically describe the properties observed in a dataset. This paper addresses the issue of efficiently extracting these summaries and rendering them to the final user, in the case where the data to be summarized are stored in a relational data base: it proposes a novel strategy that leverages the statistics about the data distribution maintained by the database system. This paper shows that reliable summaries can be very efficiently estimated based on these statistics only and without any costly data access. Additionally, it proposes a visualization of the set of extracted summaries that offers a fruitful interactive exploration tool to the user. Experiments performed on two real data bases show the relevance and efficiency of the proposed approach: with a negligible loss of accuracy, we provide the first linguistic summarization approach whose processing time does not depend on the size of the dataset. The generation of estimated linguistic summaries takes less than one second even for dataset containing millions of tuples.

In classical logic, Modus Ponens allows to infer new knowledge in the case where the antecedent of a given rule is observed, establishing that the rule conclusion then holds. Approximate reasoning extends the principle to the case where the observation does not totally match the rule antecedent. Several approaches have been proposed to deal with the extreme case where the observation is actually disjoint from the rule antecedent, using different principles to guide inference and avoid producing total uncertainty. This paper studies two of them, namely Geometric Compatibility Modification (GCM) and the Transformation-based Constraint-Guided Generalised Modus Ponens (T-CGMP), that respectively perform a type of approximate analogical reasoning and extend the GMP: it provides an indepth comparison, to determine their relationships, common points and distinct features. It thus provides guidelines for the definition of fuzzy inference schemes.

This paper studies a well-established fuzzy subspace clustering paradigm and identifies a discontinuity in the produced solutions, which assigns neighbor points to different clusters and fails to identify the expected subspaces in these situations. To alleviate this drawback, a regularization term is proposed, inspired from clustering tasks for graphs such as spectral clustering. A new cost function is introduced, and a new algorithm based on an alternate optimization algorithm, called Weighted Laplacian Fuzzy Clustering, is proposed and experimentally studied.

This paper proposes an experimental comparison between human imprecise calculation and fuzzy arithmetic: an empirical study has been conducted to collect real intervals resulting from products and additions with imprecise operands from participants. Fuzzy intervals are elicited from these data and fuzzy arithmetic is applied to the collected imprecise operands. Comparisons show that the fuzzy product and addition differ from the way human beings perform these operations. Moreover, they show that the participants, rather than taking into account the imprecisions in the calculations, realise exact calculation and in the end approximate the exact result.

The T-CGMP inference scheme combines the GMP principles with additional constraints to guide inference in the case where the observation does not match the rule premise. This paper studies its exploitation in the case where several rules are available, in particular considering the case of two rules: it proposes a local variant of T-CGMP and examines its behaviour in this interpolation-like framework, highlighting its specific features, in particular the original uncertain outputs it produces to keep track of shape mismatches. The resulting inference scheme can thus be seen as an intermediary between GMP and interpolation.

Imprecise numerical expressions, such as "about 100 meters", are pervasive in natural language. Mobile robotics, Geographic Information Systems, intelligent personal assistants as well as database querying applications are required to automatically and accurately interpret such expressions, called Approximate Numerical Expressions (ANE). The main challenge is to determine their numerical boundaries that sound plausible to users. The aim of this paper is to provide guidelines to interpret ANEs that are independent from the domain and the formal representations. We identified three arithmetical properties and examined their involvement in ANE interpretation as intervals of denoted values. The implicit assumption of symmetry of the intervals was also tested. To do so, 146 participants were asked to provide the intervals corresponding to 24 ANEs in a semantically neutral context. Results suggest that the properties of ANEs we identified are key factors in their interpretation while symmetry is not always maintained. This study contributes towards an understanding of how users process ANEs and its results can be used to improve intelligent interfaces that lead to better users' satisfaction and natural interaction between him/her and the system.

Generalised Modus Ponens (GMP) allows to perform logical inference in the case where an observation partially matches the premise of an implication, enriching the rule exploitation as compared to binary classical logic. This paper proposes to further enhance the rule exploitation, integrating additional constraints to guide the inference, both to reduce uncertainty in case of partial match and to perform inference in the case of an observation disjoint from the rule premise. These constraints are expressed as logical predicates derived from properties that characterise the observation, in an absolute way or relatively to the rule premise. An extension of GMP is proposed, to take into account the constraints, based on transformation operations applied to the fuzzy sets involved in the rule and the observation. An instantiation to a GMP preserving graduality and ambiguity is established and its validity is proven.

This paper addresses the task of community detection and proposes a local approach based on a distributed list building, where each vertex broadcasts basic information that only depends on its degree and that of its neighbours. A decentralised external process then unveils the community structure. The relevance of the proposed method is experimentally shown on both artificial and real data.

An important aspect of interpretability in Fuzzy Linguistic Summaries (FLS) is the absence of opposition therein, which is not guaranteed by the the current approaches used for their generation, possibly leading to confusion for the end-user. In this paper, we first introduce a 3-level hierarchy to organise the models of opposition starting from simpler sentences, then enriched with generalised quantifiers and thirdly considering the several negation operators allowed by fuzzy logic. We then introduce a general model of opposition for FLS sentences, which we propose to represent as a 4-dimensional cube. We additionally discuss the antonym property in this analysis framework and prove it for general protoforms.

The cross-listing mechanism enables that companies collect funds and investors invest in capital markets of foreign countries. Among the objectives are the increase in the liquidity, the reduction of the risk and of the capital cost. In this context, this paper analyses the relationship of dually listed stocks of Brazilian companies, simultaneously traded on the São Paulo stock Exchange and New York Exchange, through American Depositary Receipts (ADR). In this sense, we evaluate which of the markets has the greatest influence on the pricing of those assets. For this purpose, we extract knowledge in the form of linguistic summaries representing attribute co-variations, enriched by different types of additional information which characterize the context and describe the co-variation type.

Information cross-checking is an essential step of the trust building process that grants it its dynamics: it assesses the credibility dimension, finding confirmations or invalidations that respectively increase or weaken the current trust level of a considered piece of information and whose order influences its final value. This paper proposes a model of credibility integration that realistically takes into account even dubious confirmations and invalidations, allowing to represent a wide range of dynamic credulity stances when faced with contradictory information streams. It is formalised in an extended multivalued logic framework and illustrated with several examples to highlight the variety of behaviours it captures.

Linguistic summarization is a data mining and knowledge discovery approach to extract patterns and sum up large volume of data into simple sentences. There is a large research in generating linguistic summaries which can be used to better understand and communicate about patterns, evolution and long trends in numerical, time series or labelled data. The objective of this work is to develop a computational system capable of automatically generating linguistic descriptions of time series data of septic shock patients containing labelled data, not only of the whole series, but also on the differences between subsets of the data. This is of particular interest in septic shock, as the differences between patients are not well understood. For this purpose we propose a new type of differential summaries, based on a numerical criterion assessing the characteristics of the summary on each subset of interest. Furthermore, this paper proposes an extension of linguistic summaries to provide temporal and categorical contextualization. This is of particular interest in healthcare to detect differences related to a condition or illness as well as the effectiveness of the administered treatment.

Clustering methods are of a particular interest to discover and to summarize the structure of a data set. However, interpreting clusters may be abstruse for unexperienced users who most of the time possess their own vocabulary to describe data and properties. In this article, an approach is proposed to determine and quantify how appropriate a user-defined vocabulary is regarding the structure captured on the data distribution using a clustering method. Two measures of vocabulary appropriateness based on clustering are proposed and tested on artificial data.

Gradual itemsets of the form “the more/less A, the more/less B” extract knowledge in the form of correlations between attributes. The methods for extracting such itemsets can generate contradictory itemsets, for example simultaneously producing the itemsets “the more A, the more B” and “the more A, the less B”. To process these contradictions, we propose a constrained definition of the gradual itemset support. In particular, it does not only depend on the considered itemset, but also on its potential contradictors. An algorithm to efficiently compute the proposed global proper gradual support is defined, as well as two methods for extracting frequent gradual itemsets according to this new support definition. Experimental results obtained from a real dataset highlight the relevance of the approach.

This paper addresses the task of detecting Internet buzzes, defined as amplification phenomena, i.e. the diffusion on a very large scale of an Internet content, massively taken up within a short period of time. It proposes two approaches based on the citation graph that represents hyperlinks relation between websites. The first method detects temporal abnormalities in the number of citations of an information source, identifying information sources that undergo a surge of their direct citations. The second method exploits higher level cues, based on the definition of the dynamic cumulative visibility of an article. It captures the notion of citation cascade that is central to the specific type of buzzes related to rumour. Both detection approaches are illustrated, respectively on real data extracted from the Web and on realistic simulated data. The experimental study shows the relevance of the proposed methods and highlights their differences.

We propose an extension of the spherical K-means algorithm to deal with settings where the number of data points is largely inferior to the number of dimensions. We assume the data to lie in local and dense regions of the original space and we propose to embed each cluster into its specific ellipsoid. A new objective function is introduced, analytical solutions are derived for both the centroids and the associated ellipsoids. Furthermore, a study on the complexity of this algorithm highlights that it is of same order as the regular K-means algorithm. Results on both synthetic and real data show the efficiency of the proposed method.

Fuzzy gradual rules of the form the more X is A, the more Y is B linguistically express information about the correlation between attributes and their co-variation. They thus provide valuable information summarizing the trends observed in a given data set. In this paper, we consider strengthened fuzzy gradual rules, i.e. gradual rules enriched with a clause introduced by the expression “all the more”: such rules of the form the more X is A, the more Y is B, all the more Z is C offer additional precisions on the relation between the attributes. We study the definition of such strengthened rules, discussing their possible semantics, considering several interpretations of fuzzy gradual rules. We then propose quality criteria as well as a mining algorithm.

Given the very ambiguous and imprecise nature of sentiments and of their expressions, this survey focuses on approaches making use of components of graduality in the task of automatic sentiments analysis. To that aim, we review methods taking account of intrinsic psychological models components of graduality as well as extrinsic components issued from computational intelligence approaches. In particular, beyond psychological models of sentiments that define affective states as multidimensional vectors in affective continuous spaces, we identify three components of graduality, namely composition or blending, intensity and inheritance. In our discussion, we review how fuzzy set theory as well as other gradual structures based on a vectorial representation are employed to describe affective states as complex or imprecise entities. Finally, we focus on verbal expressions of sentiments and more specifically, we discuss the use of components of graduality in order to deal with sentiments complex and subtle expressions issued from the expressive power of natural languages.

This paper introduces a new tool for web usage mining and visualization that relies on the bio-mimetic relational clustering algorithm Leader Ant and the definition of prototypes based on typicality computation to produce an efficient visualization of the activity of users on a website. The tool is evaluated on a real web log file from the French museum of Bourges website and shows that it can easily produce meaningful visualizations of typical user navigation.

In this paper, a typicality-based clustering algorithm is proposed: it exploits typicality degrees defined in a prototype construction framework to identify a decomposition of the dataset into homogeneous and distinct clusters and to provide characteristic representatives of the obtained clusters, so as to summarise the initial dataset. The proposed algorithm can be applied both to vectorial and non vectorial data, such as trees for instance. Tests performed on artificial and real data illustrate the interest of the proposed approach.

Natural concept modelling aims at representing numerically semantic knowledge; generally, experts are asked to provide examples of linguistic terms associated with numerical data descriptions. We propose to exploit directly non labelled databases to extract the concepts that enable a semantic description of the data. Our method consists in identifying the subgroups corresponding to the concepts and then representing them as fuzzy subsets. For the identification step, we propose an algorithm based on a conjugate iterative use of the single linkage hierarchical clustering algorithm and the fuzzy c-means, that explicitly takes into account both a separability objective and a compactness aim; the description step builds membership functions as generalized Gaussians. The adequacy of the results with spontaneous descriptions is illustrated on artificial and real databases.