Monitoring the elemental composition of materials in order to detect abnormal conditions in real-time is essential for applications like manufacturing quality control, environmental monitoring, and space exploration. This is achieved using sensors that analyze the interaction of a material with electromagnetic radiation, producing spectral data streams or a sequence of instances where each represents an ordered set of wavelengths with an associated intensity. While many unsupervised anomaly detection methods exist for tabular streaming data, their applicability to spectral streams remains underexplored. To address this gap, we consider our spectra in a multivariate stream setting and benchmark the performance of state-of-the-art tabular anomaly detection methods on this data. Furthermore, we introduce OnlineBootKNN, a novel unsupervised framework that combines k-nearest neighbors with online bootstrapping and a z-score test to detect anomalies in real-time. We demonstrate the high performance and robustness of our method, as well as the efficacy of the autoencoder-based method, KitNet, on newly simulated real-world spectral datasets. In addition, we compare their efficiency against the other tested techniques. Finally, we highlight the inherent interpretability of OnlineBootKNN, which is crucial for identifying the specific wavelengths, and thus elements, responsible for a detected anomaly.

In this paper, we introduce a novel method for embedding weighted graphs, designed to capture the significance of relationships in real-world weighted graph data. Unlike existing models that treat all relationships equally, our approach incorporates edge weights directly into the embedding process, enhancing the representation of highly weighted connections. By introducing modifications to both the scoring and loss functions, our method emphasizes the importance of weighted relationships during training. Theoretical analysis shows that our model maintains efficiency and scalability while significantly improving the capacity to represent weighted graph structures. Experimental results demonstrate that our approach consistently outperforms baseline methods, including TransE, in tasks involving weighted relationships, showcasing its robustness and applicability.

Knowledge graph embedding has emerged as a fundamental technique to represent entities and relationships in knowledge graphs within low-dimensional vector spaces. Among these methods, translation-based approaches stand out by treating relations as translations from head entities to tail entities, achieving state-of-the-art results. However, the training process of these methods can be prohibitively time-consuming, especially for large knowledge graphs, posing significant challenges in practical applications. As knowledge graphs grow in size and complexity, surpassing the capacities of existing systems, there is an urgent need for scalable solutions in knowledge representation learning. These graphs, comprising millions of nodes and billions of edges, serve as powerful data structures for representing and understanding complex networks of knowledge. Translation-based models, particularly TransE, have been instrumental in encoding structured information about entities and their relationships in low-dimensional embedding spaces. However, the current implementation of TransE is constrained to single-node machines, limiting its scalability and applicability. To address these limitations, this paper proposes leveraging parallel computing technique, such as parallel TransE with graph partitioning, to enhance scalability and efficiency in knowledge graph embedding.

The task of mining gradual itemsets holds significant importance in pattern mining, particularly when working with numerical data. It involves the discovery of covariations between attributes in the form of “The more/less X,…, the more/less Y,” referred to as gradual itemsets. However, discovering these itemsets remains challenging, partly due to the exponential combinatorial search space involved in large-scale data processing. Consequently, existing algorithms for gradual itemset mining encounter difficulties, such as slow processing speeds, and occasional failures to terminate due to the overwhelming number of candidate itemsets requiring exploration. A large number of candidates is generated, but a large proportion of them turns out to be infrequent once their supports are computed. This paper introduces an approach to streamline this process by efficiently reducing the number of candidates for which support needs to be computed through the introduction of a stricter upper-bound criterion. By circumventing the costly support computation for numerous candidate itemsets, our approach exhibits efficiency in terms of speed when applied to real databases, including large-scale databases that pose challenges for existing algorithms. Furthermore, we establish a connection in terms of pattern coverage between the two principal gradualness semantics commonly employed in the literature.

International audience

In the last decades, learning over graph data has became one of the most challenging tasks in deep learning. The generally proposed Graph Neural Network (GNN) framework computes a hidden state for every node in the graph by applying nonlinear transformations to its neighborhood. Then updates the node of interest's hidden state. Nevertheless, the node features can contain discriminative information. That can get lost over GNN transformations. In this work, we present a new variant of GNN architecture where we combine node features and GNN activations to learn node representations in the graph. We conduct extensive experiments on two graph prediction tasks (node classification and link prediction) and show that our method can match and outperforms state-of-the-art results on five challenging datasets.

Heterogeneous graphs (HG) are an effective way of abstracting complex systems, including social, biological, and economic systems. However, modeling these graphs is challenging due to their high dimensionality, sparsity, and heterogeneity. Traditional approaches designed for homogeneous graphs struggle to handle the diverse types of entities and relationships present in HG, leading to a loss of information and potentially inaccurate embeddings. To address these challenges, we introduce a novel method, Trans-Trip: Translation-based embedding with Triplets for Heterogeneous Graphs, that leverages the power of triplets of (entity, relation, entity) to accurately represent the various types of relationships among nodes and links. Trans-trip effectively captures the rich semantics embedded in HGs, overcoming the challenges of global coherence and entity projection. By leveraging the flexibility and interpretability of triplets, our method can handle the multi-types nodes/links present in HG and can capture complex higher-order structures. We demonstrate the effectiveness of our proposed method on several benchmark datasets, showing that it outperforms existing embedding methods. Trans-trip provides a more accurate and interpretable representation of HG, which can be used across various fields, such as biology, social networks, and e-commerce.

One of the critical tasks of data stream analysis is anomaly detection. Various methods based on multiple assumptions have been reported in the literature. However, there is still a lack of experimental comparison of those methods, which makes it difficult to choose a specific one. In this paper, we compared unsupervised data stream abnormal point detection methods on various datasets with emphasis on their performance and runtime, as well as the presence of concept drift, seasonality, trend, and cycle as a characteristic of the dataset. Our experiments show that forecasting-based methods are the ones managing the best seasonality and trend, and lightweight models performing online gradient descent have a lower execution time. The details of our experiments are available online.

Hard and fuzzy clustering algorithms are part of the partition-based clustering family. They are widely used in real-world applications to cluster numerical and categorical data. While in hard clustering an object is assigned to a cluster with certainty, in fuzzy clustering an object can be assigned to different clusters given a membership degree. For both types of method an entropy can be incorporated into the objective function, mostly to avoid solutions raising too much uncertainties. In this paper, we present an extension of a fuzzy clustering method for categorical data using fuzzy centroids. The new algorithm, referred to as Categorical Fuzzy Entropy (CFE), integrates an entropy term in the objective function. This allows a better fuzzification of the cluster prototypes. Experiments on ten real-world data sets and statistical comparisons show that the new method can efficiently handle categorical data.

Mining frequent simultaneous attribute co-variations in numerical databases is also called frequent gradual pattern problem. Few efficient algorithms for automatically extracting such patterns have been reported in the literature. Their main difference resides in the variation semantics used. However in applications with temporal order relations, those algorithms fail to generate correct frequent gradual patterns as they do not take this temporal constraint into account in the mining process. In this paper, we propose an approach for extracting frequent gradual patterns for which the ordering of supporting objects matches the temporal order. This approach considerably reduces the number of gradual patterns within an ordered data set. The experimental results show the benefits of our approach.

Ionizing-radiation-resistant bacteria (IRRB) could be used for bioremediation of radioactive wastes and in the therapeutic industry. Limited computational works are available for the prediction of bacterial ionizing radiation resistance (IRR). In this work, we present ABClass, an in silico approach that predicts if an unknown bacterium belongs to IRRB or ionizing-radiation-sensitive bacteria (IRSB). This approach is based on a multiple instance learning (MIL) formulation of the IRR prediction problem. It takes into account the relation between semantically related instances across bags. In ABClass, a preprocessing step is performed in order to extract substructures/motifs from each set of related sequences. These motifs are then used as attributes to construct a vector representation for each set of sequences. In order to compute partial prediction results, a discriminative classifier is applied to each sequence of the unknown bag and its correspondent related sequences in the learning dataset. Finally, an aggregation method is applied to generate the final result. The algorithm provides good overall accuracy rates. ABClass can be downloaded at the following link: http://homepages.loria.fr/SAridhi/software/MIL/.

Presents corrections to the paper, A novel computational approach for global alignment for multiple biological networks,” (Djeddi, W.E., et al), Trans. Comput. Biol. Bioinf., vol. 15, no. 6, pp. 2060–2066, Nov./Dec. 2018.

Gradual patterns that capture the order correlations of the form "The more/less X, then the more/less Y" play an important role in many real world applications. In this paper, we propose an approach for extracting (closed) frequent gradual patterns when the ordering of supporting objects matches the temporal order. This approach allows to reduce the quantity of mined patterns when the objects follow a temporal order relation. The experimental results obtained on the paleoecological data show the efficiency of our approach and the interpretation of those results bring new knowledge to paleoecological experts.

Due to the rapid progress of biological networks for modeling biological systems, a lot of biomolecular networks have been producing more and more protein-protein interaction (PPI) data. Analyzing protein-protein interaction networks aims to find regions of topological and functional (dis)similarities between molecular networks of different species. The study of PPI networks has the potential to teach us as much about life process and diseases at the molecular level. Although few methods have been developed for multiple PPI network alignment and thus, new network alignment methods are of a compelling need. In this paper, we propose a novel algorithm for a global alignment of multiple protein-protein interaction networks called MAPPIN. The latter relies on information available for the proteins in the networks, such as sequence, function, and network topology. Our algorithm is perfectly designed to exploit current multi-core CPU architectures, and has been extensively tested on a real data (eight species). Our experimental results show that MAPPIN significantly outperforms NetCoffee in terms of coverage. Nevertheless, MAPPIN is handicapped by the time required to load the gene annotation file. An extensive comparison versus the pioneering PPI methods also show that MAPPIN is often efficient in terms of coverage, mean entropy, or mean normalized.

Due to its ability to solve nonlinear problems, Artificial Neural Network (ANN) could be applied in several areas of life. However, defining its architecture for solving a given problem is not formalized and remains an open research problem. On the other hand the complexity of such a technique due to its “black box” aspect, makes its interpretation more tedious. Since optimal factors completely cover the data and therefore give an explanation to these data, we propose in this paper to build feedforward ANNs using the optimal factors obtained from the boolean context representing a data. We show through experiments and comparisons on the use datasets that this approach provides relatively better results than those existing in the literature.

The huge number of association rules represent the main hamper that a decision maker faces. In order to bypass this hamper, an efficient selection of rules has to be performed. Since selection is necessarily based on evaluation, many interestingness measures have been proposed. However, the abundance of these measures gave rise to a new problem, namely the heterogeneity of the evaluation results and this created confusion to the decision. In this respect, we propose a novel approach to discover interesting association rules without favoring or excluding any measure by adopting the notion of dominance between association rules. Our approach bypasses the problem of measure heterogeneity and unveils a compromise between their evaluations. Interestingly enough, the proposed approach also avoids another non-trivial problem which is the threshold value specification.

The current development of knowledge discovery domain has pointed out a high number of applications where the need of explanation is at the heart of the process. Using neural networks for those applications requires to be able to provide a set of rules extracted from the trained neural networks, that can help the user to comprehend the learning process. The current literature reports two kinds of rules: `if condition then conclusion' (called if-then) and `if m of conditions then conclusion' (also called MofN). We propose a new method able to extract one intermediate structure (called generators list) from which it is possible to extract both forms of rules. The extracted structure is a generic representation that gives the possibility to the user to visualize each form of rules extracted from the multilayer artificial neural networks.

BACKGROUND: This paper deals with the preprocessing of protein sequences for supervised classification. Motif extraction is one way to address that task. It has been largely used to encode biological sequences into feature vectors to enable using well-known machine-learning classifiers which require this format. However, designing a suitable feature space, for a set of proteins, is not a trivial task. For this purpose, we propose a novel encoding method that uses amino-acid substitution matrices to define similarity between motifs during the extraction step. RESULTS: In order to demonstrate the efficiency of such approach, we compare several encoding methods using some machine learning classifiers. The experimental results showed that our encoding method outperforms other ones in terms of classification accuracy and number of generated attributes. We also compared the classifiers in term of accuracy. Results indicated that SVM generally outperforms the other classifiers with any encoding method. We showed that SVM, coupled with our encoding method, can be an efficient protein classification system. In addition, we studied the effect of the substitution matrices variation on the quality of our method and hence on the classification quality. We noticed that our method enables good classification accuracies with all the substitution matrices and that the variances of the obtained accuracies using various substitution matrices are slight. However, the number of generated features varies from a substitution matrix to another. Furthermore, the use of already published datasets allowed us to carry out a comparison with several related works. CONCLUSIONS: The outcomes of our comparative experiments confirm the efficiency of our encoding method to represent protein sequences in classification tasks.

In previous works, we showed how sequential pattern mining can be used to extract a partial problem space from logged user interactions for a procedural and ill-defined domain where classic domain knowledge acquisition approaches don't work well. In this paper, we describe in details how such a problem space can support important tutoring services such as (1) recognizing the plan of a learner, (2) providing hints and (3) estimating the profile of a learner including its expertise level and missing or misunderstandood skills.

The extremely large number of association rules that can be drawn from – even reasonably sized datasets, bootstrapped the development of more acute techniques or methods to reduce the size of the reported rule sets. In this context, the battery of re

Providing efficient and easy-to-use graphical tools to users is a promising challenge of data mining (DM). These tools must be able to generate explicit knowledge and to restitute it. Visualization techniques have shown to be an efficient solution to achieve such goal. Even though considered as a key step in the mining process, the visualization step of association rules received much less attention than that paid to the extraction one. Nevertheless, some graphical tools have been developed to extract and visualize association rules. In those tools, various approaches are proposed to filter the huge number of association rules before the visualization step. However both DM steps (association rule extraction and visualization) are treated separately in a one way process. Our approach differs, and uses meta-knowledge to guide the user during the mining process. Standing at the crossroads of DM and Human-Computer Interaction (HCI), we present an integrated framework covering both steps of the DM process. Furthermore, our approach can easily integrate previous techniques of association rule visualization.

Traditional framework for mining association rules has pointed out the derivation of many redundant rules, in order to be reliable in a decision making process, such discovered rules have to be concise and easily understandable for users or as well as an input to visualization tools. We present a 3D histograms-based visualization prototype for handling generic bases of association rules. An interesting feature of the prototype is that it provides a "contextual" exploration of such rule set. Such additional displayed knowledge, based on the construction of fuzzy meta-rules, enhances man-machine interaction by emulating a cooperative behavior.

Mining frequent closed itemsets is one effective method to analyse frequent pattern, and further, to generate association rules. Several algorithms were proposed to generate frequent closed itemsets, including CLOSE, A-CLOSE, CLOSET, CHARM and CLOSET + etc. However it's still hard for these algorithms to deal with dense and very large data. In this paper, we analyze the search space of frequent closed itemsets and propose a new decomposition algorithm for mining frequent closed itemsets called PFC. PFC can dynamically generate non-overlapping partitions of the search space and mine frequent closed itemsets in each partition. Furthermore, each partition is independent and only shares the same source data with other partitions. So it is possible to implement PFC with multi-threads or parallel methods, and prune efficiently the search space of frequent closed itemsets. In this study, P FC is implemented in Java. We compare PFC with an author's C++ version of CLOSET + on some large VCI repository datasets and on the worst case. The preliminary experimental results demonstrate good performance of PFC for dealing with dense and very large data.

Traditional framework for mining association rules has pointed out the derivation of many redundant rules. In order to be reliable in a decision making process, such discovered rules have to be both concise and easily understandable for users, and/or as an input to visualization tools [P. Adriaans et al. (1997)]. We present a graphical visualization prototype for handling generic bases of association rules. We discuss also the most adequate graphical visualization technique depending on the intrinsic structure of the generic bases of association rules. An interesting feature of the prototype is that it provides a "contextual" exploration of such rule set. Such exploration, based on the discovery of fuzzy meta-rules, enhances man-machine interaction by emulating a cooperative behavior.

Concept lattice is an effective tool and platform for data analysis and knowledge discovery such as classification or association rules mining. The lattice algorithm to build formal concepts and concept lattice plays an essential role in the application of concept lattice. We propose a new efficient scalable lattice-based algorithm: ScalingNextClosure to decompose the search space of any huge data in some partitions, and then generate independently concepts (or closed itemsets) in each partition. The experimental results show the efficiency of this algorithm.

(2002). Concept lattice-based knowledge discovery in databases. Journal of Experimental & Theoretical Artificial Intelligence: Vol. 14, No. 2-3, pp. 75-79.

Multistrategy learning (MSL) consists of combining at least two different learning strategies to bring out a powerful system, where the drawbacks of the basic algorithms are avoided. In this scope, instance-based learning (IBL) techniques are often used as the basic component. However, one of the major drawbacks of IBL is the prototype selection problem which consists in selecting a subset of representative instances in order to reduce the classification process. This paper presents a novel approach which consists of three steps. The first one builds a set of lattice-based hypotheses that characterize the training data set. Given an unseen example, the second step selects a subset of training instances through the way they verify the same hypotheses as the unseen example. Finally the last step uses this subset of training instances as the prototypes for the classification of the unseen example. Results of experiments that we conducted show the effectiveness of our approach compared to standard ML techniques on different datasets.

Concept learning is one of the most studied areas in machine learning. A lot of work in this domain deals with decision trees. In this paper, we are concerned with a different kind of technique based on Galois lattices or concept lattices. We present a new semilattice based system, IGLUE, that uses the entropy function with a tap-down approach to select concepts during the lattice construction. Then IGLUE generates new relevant numerical features by transforming initial boolean features over these concepts. IGLUE uses the new features to redescribe examples. Finally, IGLUE applies the Mahanalobe's distance as a similarity measure between examples.

The previously-reported LEGAL system is an empirical machine learning system based on Galois Lattice. Its aim is first to produce a semi-lattice from a concept denoted by a set of objects which are described with binary attributes. Then using some selected attribute conjunctions in the semi-lattice and a majority vote principle, LEGAL predicts new examples from unseen objects. This paper describes a new version LEGAL-E and its application to two biological problems: the prediction of splice junctions sites and the promoter recognition. Results obtained are far better than those of some symbolic learning systems, and are as better as those of some best neural networks methods. Moreover some empirical properties shared by LEGAL-E and neural networks are described. Finally this paper shows how the semi-lattice can be used as a dynamic neural network architecture in order to combine both learning techniques for knowledge refinement.>