Machine Learning (ML) has become an essential tool for modeling complex phenomena, offering robust predictions and comprehensive data analysis. Nevertheless, the lack of interpretability in these predictions often results in a closed-box effect, which the field of eXplainable Machine Learning (XML) aims to address. Local attributive XML methods, in particular, provide explanations by quantifying the contribution of each attribute to individual predictions, referred to as influences. This type of explanation is the most acute as it focuses on each instance of the dataset and allows the detection of individual differences. Additionally, aggregating local explanations allows for a deeper analysis of the underlying data. In this context, influences can be considered as a new data space to reveal and understand complex data patterns. We hypothesize that these influences, derived from ML explanations, are more informative than the original raw data, especially for identifying homogeneous groups within the data. To identify such groups effectively, we utilize a clustering approach. We compare clusters formed using raw data against those formed using influences computed by various local attributive XML methods. Our findings reveal that clusters based on influences consistently outperform those based on raw data, even when using models with low accuracy.

When applying clustering techniques to data exploration, ensuring the practical usefulness of clusters by aligning them with expert knowledge is highly desirable. A recent approach, known as supervised clustering, addresses this by selecting a target feature and constructing a Target Explanation Space (TES) using a supervised model combined with local feature attribution methods such as LIME or SHAP. While TES enhances clustering performance, its lack of interpretability remains a significant limitation. To address this, we introduce the first hierarchical supervised clustering pipeline that generates interpretable, nonoverlapping rules directly in the original data space-while still lever-aging the improved clustering achieved in TES. Experimental results demonstrate that the rules produced are not only more concise but also more comprehensive than those from existing methods, enabling experts to effectively balance interpretability and predictive accuracy.

While scoring nodes in graphs to understand their importance (e.g., in terms of centrality) has been investigated for decades, comparing nodes in property graphs based on their properties has not, to our knowledge, yet been addressed. In this paper, we propose an approach to automatically extract comparison of nodes in property graphs, to support the interactive exploratory analysis of said graphs. We first present a way of devising comparison indicators using the context of nodes to be compared. Then, we formally define the problem of using these indicators to group the nodes so that the comparisons extracted are both significant and not straightforward. We propose various heuristics for solving this problem. Our tests on real property graph databases show that simple heuristics can be used to obtain insights within minutes while slower heuristics are needed to obtain insights of higher quality.

Exploratory Data Analysis (EDA) is the tedious activity of interactively analyzing a dataset to extract insights. Many approaches aiming at supporting EDA were recently proposed. They all rely on interestingness measures to score the importance of insights. This paper surveys and categorizes the different interestingness measures proposed in the literature for approaches aiming at automating EDA. The lessons learned from this survey allow to point out promising research directions.

Exploratory Data Analysis (EDA) is the interactive process of gaining insights from a dataset. Comparisons are popular insights that can be specified with comparison queries, i.e., specifications of the comparison of subsets of data. In this work, we consider the problem of automatically computing sequences of comparison queries that are coherent, significant and whose overall cost is bounded. Such an automation is usually done by either generating all insights and solving a multi-criteria optimization problem, or using reinforcement learning. In the first case, a large search space has to be explored using exponential algorithms or dedicated heuristics. In the second case, a dataset-specific, time and energy-consuming training, is necessary. We contribute with a novel approach, consisting of decomposing the optimization problem in two: the original problem, that is solved over a smaller search space, and a new problem of generating comparison queries, aiming at generating only queries improving existing solutions of the first problem. This allows to explore only a portion of the search space, without resorting to reinforcement learning. We show that this approach is effective, in that it finds good solutions to the original multi-criteria optimization problem, and efficient, allowing to generate sequences of comparisons in reasonable time.

Recommendersystemsarenotoriouslycomplexsystemsforwhichprovidingalocalexplanationonwhyacertainitemisproposedtoaspecificuserisstillachallengingtask .Mostexplanation approachesfocusonpredictingtheratingofitemstherebymin-imizingsomediscrepancywiththerealratingsbymeansof traditionallossfunctions(e.g.,sumofsquareserror).However, mostofthetimes,ratingsmaynotfullyembraceuserpreferences concerningtherankingofitems.Tobetterembraceuserpref-erences,methodsbasedonrankinglosseshavebeenproposed eithertorecommendortoexplainwhyanitemisrecommended. Although effective at identifying the most prominent items, these methods fail to capture a realistic value for the rating attached to their explanation. This loss attached to the semantic of the recommendation can in turn arm the trust of a user in the explanation. In this paper, we propose and discuss experimental results of a simple yet effective novel loss schema that balances ranking and rating losses to provide a best of both world explanation.

We consider the problem of generating SQL notebooks of comparison queries for Exploratory Data Analysis (EDA). A comparison query allows to find insights in a dataset by specifying the comparison of subsets of data. In this paper, we study the problem of generating sequences of comparison queries that are insightful and coherent. We propose exact and approximate resolution approaches, and study their efficiency and effectiveness on artificial and real datasets, as well as with a user study.

In this short paper, we present our preliminary results for generating personalized data narrations by extracting messages from a collection of Exploratory Data Analysis (EDA) notebooks over a given dataset. The approach consists of extracting features from notebooks to learn what interesting messages they expose. Based on those interesting messages, we formalize the problem of producing a user-tailored data narration, i.e., a coherent sequence of messages matching a given user profile. We developed a proof of concept and experimented with Kaggle.com notebooks.

Post-hoc explanation aims at defining a simple local surrogate model to shed light on a prediction produced by a complex, generally black-box, model. In the general context of classification, it has been shown that local surrogates may not always be able to capture a local explanation, i.e. for a specific instance prediction, but rather depict more of a general behavior of the black-box. This problem is even more complex in a recommendation scenario where classes and decision boundaries are not explicitly defined and where data are very sparse by nature. We show in this paper that it is possible to tackle these problems with an efficient sampling around the recommendation instance to explain, to finally learn a proper local surrogate model. To this aim, this paper introduces several new approaches to capture efficiently local explanation models in the context of recommendation, all defined around a locality sample. Noticeably, and novel to this work, we show that it is possible to achieve a simple, yet better quality explanation model by not directly considering ratings, but rather implicit preferences as expressed by comparisons of pairs of ratings. We introduce to this extent a novel explainable model based on a pairwise loss RankNet architecture. Extensive experiments show that our methods can be better than state-of-the-art methods depending on the locality of the black-box model, and are much more efficient to retrieve meaningful explainable features locally.

The study of semantic aspects of human behavior is an hot topic. Most of the time, semantic sequences describe these complex behaviors. Indeed, sequences include several information as type of human activities or places. To study these complex data, we need to define new similarity measures and select appropriate clustering processes. This article proposes a semantic similarity measure, based on ontologies, which manages complex semantic elements with different levels of detail and incertitude. An application of this approach from the domain of touristic mobility shows the interest of this process.

Post-hoc explanation aims at defining a simple local surrogate model to shed light on a prediction produced by a complex, generally black-box, model. In the general context of classification, it has been shown that local surrogatesmay not be able to always capture a local explanation, i.e. for a specific instance prediction, but rather traduce more of a general behavior of the black-box. This problem is even more complex in a recommendation scenario where classes and decision boundaries are not explicitly defined and where data are very sparse by nature. We show in this paper that it is possible to tackle these problems with an efficient sampling around the recommendation instance to explain, to finally learn a proper local surrogate model. Our experiments show that our method is as accurate or better than the methods of the literature while retrieving more meaningful explainable features locally.

This paper addresses the problem of defining a subjective interestingness measure for BI exploration. Such a measure involves prior modeling of the belief of the user. The complexity of this problem lies in the impossibility to ask the user about the degree of belief in each element composing their knowledge prior to the writing of a query. To this aim, we propose to automatically infer this user belief based on the user's past interactions over a data cube, the cube schema and other users past activities. We express the belief under the form of a probability distribution over all the query parts potentially accessible to the user, and use a random walk to learn this distribution. This belief is then used to define a first Subjective Interestingness measure over multidimensional queries. Experiments conducted on simulated and real explorations show how this new subjective interestingness measure relates to prototypical and real user behaviors, and that query parts offer a reasonable proxy to infer user belief.