Abstract Transparency is a non-functional requirement of machine learning that promotes interpretable models and easily explainable outcomes. Unfortunately, interpretable classification models, such as linear, rule-based, and decision tree models, are superseded by more accurate but complex learning paradigms, such as deep neural networks and ensemble methods. More specifically, for tabular data classification, models based on tree ensembles, such as random forest or XGBoost, are still competitive compared to deep learning ones and are often preferred to the latter. However, they share the same interpretability issues, due to the complexity of the learned model and, consequently, offer low explainability of predictions. Existing solutions consist of computing some feature importance score or extracting an approximate surrogate model from the learned tree ensemble. However, these methods lead to surrogate models with either poor fidelity or questionable comprehensibility. In this paper, we propose to improve this trade-off using Goodman-Kruskal’s association measure to find groups of instances with predictions that are explained by shared groups of features. To build this structure, instances are first described by SHAP values, which capture local feature importance, and then co-clustered with features on the basis of these SHAP values. Next, a surrogate model is built as a set of shallow decision trees learned for the different groups of instances and subsets of relevant features. Our experiments show that our method produces surrogate models that explain random forest and XGBoost classifiers with competitive fidelity and higher comprehensibility compared to recent state-of-the-art competitors.

Abstract Subspace clustering is a data‐mining task that groups similar data objects and at the same time searches the subspaces where similarities appear. For this reason, subspace clustering is recognized as more general and complicated than standard clustering. In this article, we present ChameleoClust+, a bioinspired evolutionary subspace clustering algorithm that takes advantage of an evolvable genome structure to detect various numbers of clusters located in different subspaces. ChameleoClust+ incorporates several biolike features such as a variable genome length, both functional and nonfunctional elements, and mutation operators including large rearrangements. It was assessed and compared with the state‐of‐the‐art methods on a reference benchmark using both real‐world and synthetic data sets. Although other algorithms may need complex parameter settings, ChameleoClust+ needs to set only one subspace clustering ad hoc and intuitive parameter: the maximal number of clusters. The remaining parameters of ChameleoClust+ are related to the evolution strategy (eg, population size, mutation rate), and a single setting for all of them turned out to be effective for all the benchmark data sets. A sensitivity analysis has also been carried out to study the impact of each parameter on the subspace clustering quality.

Dynamic systems such as glaciers can be studied using Displacement Field Time Series (DFTS), often derived from Satellite Image Time Series. Even if data mining patterns expressing interesting displacement evolutions can be extracted from DFTS, confidence measures coming along with these series have to be considered to focus on reliable evolutions. This paper introduces a new approach for selecting displacement evolutions that are reliable, informative and complementary. Reported experiments exhibit consistent displacement evolutions of Alpine glaciers and complete the current knowledge of the area.

In this paper we present an evolutionary algorithm to tackle the subspace clustering problem. Subspace clustering is recognized as more difficult than standard clustering since it requires to identify not only the clusters but also the various subspaces where the clusters hold. We propose to tackle this problem with a bio-inspired algorithm that includes many bio-like features like variable genome length and organization, functional and non-functional elements, and variation operators including chromosomal rearrangements. These features give the algorithm a large degree of freedom to achieve subspace clustering with satisfying results on a reference benchmark with respect to state of the art methods. One of the main advantages of the approach is that it needs only one subspace clustering ad-hoc parameter: the maximal number of clusters. This is a single and intuitive parameter that sets the maximal level of details of the clustering, while other algorithms require more complicated parameter space exploration. The other parameters of the algorithm are related to the evolution strategy (population size, mutation rate, ...) and for them we use a single setting that turns out to be effective on all the datasets of the benchmark.

In this paper we present a method to summarize a satellite image time series. This summary is a small set of maps depicting salient phenomena occurring in the series over space and time. The approach is composed of a first step of extraction of spatiotemporal patterns, followed by an iterative ranking of these patterns using a swap randomization technique and a ranking based on a normalized mutual information measure. The best ranked patterns in the earliest iterations are in some sense the most informative and are used to build the summary. We present results showing that the approach is effective on both optical and radar data.

This paper presents a data mining approach for describing Satellite Image Time Series (SITS) spatially and temporally. It relies on pixel-based evolution and sub-evolution extraction. These evolutions, namely the {frequent grouped sequential patterns}, are required to cover a minimum surface and to affect pixels that are sufficiently connected. These spatial constraints are actively used to face large data volumes and to select evolutions making sense for end-users. In this paper, a specific application to fully polarimetric SAR image time series is presented. Experiments performed on a RADARSAT-2 SITS covering the Chamonix Mont Blanc test-site are used to illustrate the proposed approach.

An important aspect of satellite image time series is the simultaneous access to spatial and temporal information. Various tools allow end users to interpret these data without having to browse the whole data set. In this paper, we intend to extract, in an unsupervised way, temporal evolutions at the pixel level and select those covering at least a minimum surface and having a high connectivity measure. To manage the huge amount of data and the large number of potential temporal evolutions, a new approach based on data-mining techniques is presented. We have developed a frequent sequential pattern extraction method adapted to that spatiotemporal context. A successful application to crop monitoring involving optical data is described. Another application to crustal deformation monitoring using synthetic aperture radar images gives an indication about the generic nature of the proposed approach.

This paper presents an original data mining approach for extracting pixel evolutions and sub-evolutions from Satellite Image Time Series. These patterns, called frequent grouped sequential patterns, represent the (sub-)evolutions of pixels over time, and have to satisfy two constraints: firstly to correspond to at least a given minimum surface and secondly to be shared by pixels that are sufficiently connected. These spatial constraints are actively used to face large data volumes and to select evolutions making sense for end-users. Successful experiments on an optical and a radar SITS are presented.

In this paper, we propose an original computational approach to assist knowledge discovery in complex biological networks. First, we present an integrated model of the evolution of regulation networks that can be used to uncover organization principl

In this paper, we present a framework for automatic generation of CHR solvers given the logical specification of the constraints. This approach takes advantage of the power of tabled resolution for constraint logic programming, in order to check the validity of the rules. Compared to previous work (Apt and Monfroy 1999; Ringeissen and Monfroy 2000; Abdennadher and Rigotti 2000; Abdennadher and Rigotti 2001a), where different methods for automatic generation of constraint solvers have been proposed, our approach enables the generation of more expressive rules (even recursive and splitting rules) that can be used directly as CHR solvers.

A general approach to implement propagation and simplification of constraints consists of applying rules over these constraints. However, a difficulty that arises frequently when writing a constraint solver is to determine the constraint propagation algorithm. In this article, we propose a method for generating propagation and simplification rules for constraints over finite domains defined extensionally by, for example, a truth table or their tuples. The generation of rules is performed in two steps. First, propagation rules are generated. Propagation rules do not rewrite constraints but add new ones. Thus, the constraint store may contain superfluous constraints. Removing these constraints not only allows saving of space but also decreases the cost of constraint solving. Constraints can be removed using simplification rules. Thus, in a second step, some propagation rules are transformed into simplification rules.Furthermore, we show that our approach performs well on various examples, including Boolean constraints, multivalued logic, and Allen's qualitative approach to temporal logic. Moreover, an application taken from the field of digital circuit design shows that our approach is of practical use.

A general approach to implement propagation and simplification of constraints consists of applying rules over these constraints. However, a difficulty that arises frequently when writing a constraint solver is to determine the constraint propagation algorithm. In previous work, different methods for automatic generation of propagation rules for constraints defined over finite domains have been proposed. We present a method for generating propagation rules for constraint predicates defined by means of a constraint logic program.

Given a large set of data, a common data mining problem is to extract the frequent patterns occurring in this set. The idea presented in this paper is to extract a condensed representation of the frequent patterns called disjunction-free sets, instead of extracting the whole frequent pattern collection. We show that this condensed representation can be used to regenerate all frequent patterns and their exact frequencies.

A general approach to implement propagation and simplification of constraints consists of applying rules over these constraints. Recently, several proposals have been made to solve finite constraint satisfaction problems by means of automatically generated propagation rules [4, 9, 3]. Since propagation rules do not rewrite constraints but add new ones, the constraint store may contain superfluous constraints. Removing these constraints not only allows saving of space but also decreases the cost of constraint solving. Constraints can be removed using simplification rules. In this paper, we present a method to transform propagation rules into simplification rules using the concept of confluence. The generated rules are implemented in the language Constraint Handling Rules (CHR). An example tak en from the field of digital circuit design shows that our approach is of practical use.

We study KDD (Knowledge Discovery in Databases) processes on multidimensional data from a query point of view. Focusing on association rule mining, we consider typical queries to cope with the pre-processing of multidimensional data and the post-processing of the discovered patterns as well. We use a model and a rule-based language stemming from the OLAP multidimensional representation, and demonstrate that such a language fits well for writing KDD queries on multidimensional data. Using an homogeneous data model and our language for expressing queries at every phase of the process appears as a valuable step towards a better understanding of interactivity during the whole process.