Search

International audience; Digital data plays a central role in sciences, journalism, environment, digital humanities, etc. Open Data sharing initiatives lead to many large, interesting datasets being shared online. Some of these are RDF graphs, but other formats like CSV, relational, property graphs, JSON or XML documents are also frequent.Practitioners need to understand a dataset to decide whether it is suited to their needs. Datasets may come with a schema and/or may be summarized, however the first is not always provided and the latter is often too technical for non-IT users. To overcome these limitations, we present an end-to-end dataset abstraction approach, which (i) applies on any (semi)structured data model; (ii) computes a description meant for human users, in the form of an Entity- Relationship diagram; (ii) integrates Information Extraction and data profiling to classify dataset entities among a large set of intelligible categories. We implemented our approach in a system called Abstra, and detail its performance on various datasets.

International audience

International audience; Graphs, and notably RDF graphs, are a prominent way of sharing data. As data usage democratizes, users need help figuring out the useful content of a graph dataset. In particular, journalists with whom we collaborate are interested in identifying, in a graph, the connections between entities, e.g., people, organizations, emails, etc.We present a novel, interactive method for exploring data graphs through their data paths connecting Named Entities (NEs, in short); each data path leads to a tabular-looking set of results. NEs are extracted from the data through dedicated Information Extraction modules. Our method builds upon the pre-existing ConnectionLens platform and follow-up work on dataset abstraction. The contribution of the present work is in the interactive and efficient approach to enumerate and compute NE paths, based on an algorithm which automatically recommends subpaths to materialize, and rewrites the path queries using these subpaths. Our experiments demonstrate the interest of NE paths and the efficiency of our method for computing them.

International audience; Graphs, and notably RDF graphs, are a prominent way of sharing data. As data usage democratizes, users need help figuring out the useful content of a graph dataset. In particular, journalists with whom we collaborate [3] are interested in identifying, in a graph, the connections between entities, e.g., people, organizations, emails, etc. We propose PathWays, an interactive tool for exploring data graphs through their data paths connecting Named Entities (NEs, in short); each data path leads to a tabular-looking set of results. NEs are extracted from the data through dedicated Information Extraction modules. PathWays leverages the ConnectionLens platform [4,6] and follow-up work on dataset abstraction [5]. Its novelty lies in its interactive and efficient approach to enumerate, compute, and analyze NE paths.

International audience; Exploring Knowledge Graphs (KGs, in short) to discover facts and links is tedious even for experts with knowledge of SPARQL due to their unfamiliarity with the structure and labels of entities, classes and relations. Some KG applications require finding the connections between groups of nodes, even if users ignore the shape of these connections. However, SPARQL only allows checking if paths exist, not returning them. A recent property graph query language, GPML, allows also returning connecting paths, but not connections between three or more nodes. We propose to demonstrate RelSearch, a system supporting extended SPARQL queries, featuring standard Basic Graph Patterns (BGPs) as well as novel Connecting Tree Patterns (CTPs); each CTP requests the connections (paths, or trees) between nodes bound to variables. RelSearch evaluates such extended queries using novel algorithms which, unlike prior keyword search methods, return connections regardless of the edge directions and are independent of how we measure the quality (score) of each connection. We will demonstrate RelSearch's expressivity and efficiency using a variety of RDF graphs, user-selected score functions, and search exploration orders.

International audience; Several targeted advertising platforms offer transparency mechanisms, but researchers and civil societies repeatedly showed that those have major limitations. In this paper, we propose a collaborative ad transparency method to infer, without the cooperation of ad platforms, the targeting parameters used by advertisers to target their ads. Our idea is to ask users to donate data about their attributes and the ads they receive and to use this data to infer the targeting attributes of an ad campaign. We propose a Maximum Likelihood Estimator based on a simplified Bernoulli ad delivery model. We first test our inference method through controlled ad experiments on Facebook. Then, to further investigate the potential and limitations of collaborative ad transparency, we propose a simulation framework that allows varying key parameters. We validate that our framework gives accuracies consistent with real-world observations such that the insights from our simulations are transferable to the real world. We then perform an extensive simulation study for ad campaigns that target a combination of two attributes. Our results show that we can obtain good accuracy whenever at least ten monitored users receive an ad. This usually requires a few thousand monitored users, regardless of population size. Our simulation framework is based on a new method to generate a synthetic population with statistical properties resembling the actual population, which may be of independent interest.

International audience; Open Information Extraction (OIE) is the task of extracting tuples of the form (subject, predicate, object), without any knowledge of the type and lexical form of the predicate, the subject, or the object. In this work, we focus on improving OIE quality by exploiting domain knowledge about the subject and object. More precisely, knowing that the subjects and objects in sentences are often named entities, we explore how to inject constraints in the extraction through constrained inference and constraint-aware training. Our work leverages the state-of-the-art OpenIE6 platform, which we adapt to our setting. Through a carefully constructed training dataset and constrained training, we obtain a 29.17% F1-score improvement in the CaRB metric and a 24.37% F1-score improvement in the WIRe57 metric. Our technique has important applications-one of them is investigative journalism, where automatically extracting conflict-of-interest between scientists and funding organizations helps understand the type of relations companies engage with the scientists. Our code and data are available at https: //github.com/prajnaupadhyay/ openie-with-entities

Online political advertising has become the cornerstone of political campaigns. The budget spent solely on political advertising in the U.S. has increased by more than 100% from \$700 million during the 2017-2018 U.S. election cycle to \$1.6 billion during the 2020 U.S. presidential elections. Naturally, the capacity offered by online platforms to micro-target ads with political content has been worrying lawmakers, journalists, and online platforms, especially after the 2016 U.S. presidential election, where Cambridge Analytica has targeted voters with political ads congruent with their personality To curb such risks, both online platforms and regulators (through the DSA act proposed by the European Commission) have agreed that researchers, journalists, and civil society need to be able to scrutinize the political ads running on large online platforms. Consequently, online platforms such as Meta and Google have implemented Ad Libraries that contain information about all political ads running on their platforms. This is the first step on a long path. Due to the volume of available data, it is impossible to go through these ads manually, and we now need automated methods and tools to assist in the scrutiny of political ads. In this paper, we focus on political ads that are related to policy. Understanding which policies politicians or organizations promote and to whom is essential in determining dishonest representations. This paper proposes automated methods based on pre-trained models to classify ads in 14 main policy groups identified by the Comparative Agenda Project (CAP). We discuss several inherent challenges that arise. Finally, we analyze policy-related ads featured on Meta platforms during the 2022 French presidential elections period., Proceedings of the ACM Web Conference 2023 (WWW '23), May 1--5, 2023, Austin, TX, USA

National audience; Graph data management and querying has many practical applications. When graphs are very heterogeneous and/or users are unfamiliar with their structure, they may need to find how two or more groups of nodes are connected in a graph, even when users are not able to describe the connections. This is only partially supported by existing query languages, which allow searching for paths, but not for trees connecting three or more node groups. The latter is related to the NP-hard Group Steiner Tree problem, and has been previously considered for keyword search in databases. In this work, we formally show how to integrate connecting tree patterns (CTPs, in short) within a graph query language such as SPARQL or Cypher, leading to an Extended Query Language (or EQL, in short). We then study a set of algorithms for evaluating CTPs; we generalize prior keyword search work, most importantly by (i) considering bidirectional edge traversal and (ii) allowing users to select any score function for ranking CTP results. To cope with very large search spaces, we propose an efficient pruning technique and formally establish a large set of cases where our algorithm, MOLESP, is complete even with pruning. Our experiments validate the performance of our CTP and EQL evaluation algorithms on a large set of synthetic and real-world workloads.

International audience

International audience; Fact-checking is a staple of journalists’ work. As more and more important data is available in electronic format, computational fact-checking, leveraging digital data sources, has been gaining interest from the journalists as well as the computer science community. A particular class of interesting data sources are statistics, that is, numerical data compiled mostly by governments, administrations, and international organizations. We propose to demonstrate StatCheck, a fact-checking system specialized in the French media arena. StatCheck builds upon a prior pipeline [CMT17 , CMT18 , DCMT19] for fact-checking statistical claims against the INSEE national statistic institute’s data and reports. In collaboration with factchecking journalists from France Info (Radio France), we have revisited and improved this pipelined, and enlarged its database by an order of magnitude by adding all Eurostat statistics. StatCheck also includes two novel statistic claim extraction modules, generalizing and improving over [DCMT19]. Based on the journalists’ feedback, we built a new user interface, suiting better their needs. We will showcase StatCheck on a variety of scenarios, where statistical claims made in social media are checked against our statistic data.

International audience; Digital data sharing leads to unprecedented opportunities to develop data-driven systems for supporting economic activities, the social and political life, and science. Many open-access datasets are RDF (Linked Data) graphs, but others are JSON or XML documents, CSV files, Neo4J property graphs, etc. Potential users need to understand a dataset in order to decide if it is useful for their goal. While some published datasets come with a schema and/or documentation, this is not always the case. We demonstrate Abstra, a dataset abstraction system, which applies on a large variety of data models. Abstra computes a description meant for humans, and integrates Information Extraction to classify dataset content among a set of categories of interest to the user. Our abstractions are conceptually close to Entity-Relationship diagrams, but our entities can have deeply nested structure.

International audience

International audience; Access to high-quality data repositories and benchmarks have been instrumental in advancing the state of the art in many experimental research domains. While advanced analytics tasks over time series data have been gaining lots of attention, lack of such community resources severely limits scientific progress. In this paper, we present Exathlon, the first comprehensive public benchmark for explainable anomaly detection over high-dimensional time series data. Exathlon has been systematically constructed based on real data traces from repeated executions of large-scale stream processing jobs on an Apache Spark cluster. Some of these executions were intentionally disturbed by introducing instances of six different types of anomalous events (e.g., misbehaving inputs, resource contention, process failures). For each of the anomaly instances, ground truth labels for the root cause interval as well as those for the extended effect interval are provided, supporting the development and evaluation of a wide range of anomaly detection (AD) and explanation discovery (ED) tasks. We demonstrate the practical utility of Exathlon's dataset, evaluation methodology, and end-toend data science pipeline design through an experimental study with three state-of-the-art AD and ED techniques.

International audience; Big data processing at the production scale presents a highly complex environment for resource optimization (RO), a problem crucial for meeting performance goals and budgetary constraints of analytical users. The RO problem is challenging because it involves a set of decisions (the partition count, placement of parallel instances on machines, and resource allocation to each instance), requires multi-objective optimization (MOO), and is compounded by the scale and complexity of big data systems while having to meet stringent time constraints for scheduling. This paper presents a MaxCompute based integrated system to support multi-objective resource optimization via ne-grained instance-level modeling and optimization. We propose a new architecture that breaks RO into a series of simpler problems, new ne-grained predictive models, and novel optimization methods that exploit these models to make effective instance-level RO decisions well under a second. Evaluation using production workloads shows that our new RO system could reduce 37-72% latency and 43-78% cost at the same time, compared to the current optimizer and scheduler, while running in 0.02-0.23s.

International audience; Freedom of the press is under threat worldwide, and the quality of information that people have access to is dangerously degraded, under the joint threat of non-democratic governments and fake information propagation. The press as an industry needs powerful data management tools to help them interpret the complex reality surrounding us. Since 2018, I have been cooperating with journalists from Le Monde, France's leading newspaper, in devising tools for analyzing large and heterogeneuos data sources that they are interested in. This research has been embodied in ConnectionLens, a graph ETL tool capable of ingesting heterogeneous data sources into a graph, enriched (with the help of ML methods) with entities extracted from data of any type. On such integrated graphs, we devised novel algorithms for keyword search, and combine them in more recent research with structured querying. The talk describes the architecture and main algorithmic challenges in building and exploiting ConnectionLens graphs, illustrated in particular on an application where we study conflicts of interest in the biomedical domain. This is joint work with A. Anadiotis, O. Balalau, H. Galhardas and many others. ConnectionLens Web site (papers+code): https://team.inria.fr/cedar/connectionlens/. This research has been funded by Agence Nationale de la Recherche AI Chair SourcesSay (https://sourcessay.inria.fr).

International audience; Citations are crucial in scientific works as they help position a new publication. Each citation carries a particular intent, for example, to highlight the importance of a problem or to compare against results provided by another method. The authors' intent when making a new citation has been studied to understand the evolution of a field over time or to make recommendations for further citations. In this work, we address the task of citation intent prediction from a new perspective. In addition to textual clues present in the citation phrase, we also consider the citation graph, leveraging high-level information of citation patterns. In this novel setting, we perform a thorough experimental evaluation of graph-based models for intent prediction. We show that our model, GraphCite, improves significantly upon models that take into consideration only the citation phrase. Our code is available online.

Online Transaction Processing (OLTP) deployments are migrating from on-premise to cloud settings in order to exploit the elasticity of cloud infrastructure which allows them to adapt to workload variations. However, cloud adaptation comes at the cost of redesigning the engine, which has led to the introduction of several, new, cloud-based transaction processing systems mainly focusing on: (i) the transaction coordination protocol, (ii) the data partitioning strategy, and, (iii) the resource isolation across multiple tenants. As a result, standalone OLTP engines cannot be easily deployed with an elastic setting in the cloud and they need to migrate to another, specialized deployment., In this paper, we focus on workload variations that can be addressed by modern multi-socket, multi-core servers and we present a system design for providing fine-grained elasticity to multi-tenant, scale-up OLTP deployments. We introduce novel components to the virtualization software stack that enable on-demand addition and removal of computing and memory resources. We provide a bi-directional, low-overhead communication stack between the virtual machine and the hypervisor, which allows the former to adapt to variations coming both from the workload and the resource availability. We show that our system achieves NUMA-aware, millisecond-level, stateful and fine-grained elasticity, while it is not intrusive to the design of state-of-the-art, in-memory OLTP engines. We evaluate our system through novel use cases demonstrating that scale-up elasticity increases resource utilization, while allowing tenants to pay for actual use of resources and not just their reservation.

International audience; The growth of online learning, enabled by the availability on the Internet of different forms of didactic materials such as MOOCs and Intelligent Tutoring Systems (ITS), in turn, increases the relevance of personalized instructions for students in an adaptive learning environment. There are increasing interests as well as many challenges in the application of Artificial Intelligence (AI) techniques in educational settings to provide adaptive learning content to learners. Knowledge assessment is necessary for providing an adaptive learning environment. A student model serves as a fundamental building block of knowledge assessment in an adaptive learning environment. This paper intends to review the development of dominant families of student models with psychometric theory in early educational research, recent adaptations, and advances with machine learning and deep learning techniques. Our review covers not only the important families of student models but also why they were invented from both theoretical and practical viewpoints with AI and educational perspectives. We believe that the discussion covered in this review will be a valuable reference of introductory insights to AI for educational researchers, as well as an endeavor of introducing basic psychometric perspectives to AI experts for knowledge assessment in the field of learning science. Finally, we provide recent challenges and some potential directions for developing efficient knowledge assessment techniques in future adaptive learning ecosystems.

International audience; Institutions collect massive learning traces but they may not disclose it for privacy issues. Synthetic data generation opens new opportunities for research in education. In this paper we present a generative model for educational data that can preserve the privacy of participants, and an evaluation framework for comparing synthetic data generators. We show how naive pseudonymization can lead to re-identification threats and suggest techniques to guarantee privacy. We evaluate our method on existing massive educational open datasets.

International audience; To strengthen public trust and counter disinformation, computational fact-checking, leveraging digital data sources, attracts interest from the journalists and the computer science community. A particular class of interesting data sources comprises statistics, that is, numerical data compiled mostly by governments, administrations, and international organizations. Statistics are often multidimensional datasets, where multiple dimensions characterize one value, and the dimensions may be organized in hierarchies. This paper describes STATCHECK, a statistic fact-checking system jointly developed by the authors, which are either computer science researchers or fact-checking journalists working for a French-language media with a daily audience of more than 15 millions (aud, 2022). The technical novelty of STATCHECK is twofold: (i) we focus on multidimensional, complex-structure statistics, which have received little attention so far, despite their practical importance; and (ii) novel statistical claim extraction modules for French, an area where few resources exist. We validate the efficiency and quality of our system onlarge statistic datasets (hundreds of millions of facts), including the complete INSEE (French)and Eurostat (European Union) datasets, as well as French presidential election debates.

International audience; Digital data is produced in many data models, ranging from highly structured (typically relational) to semi-structured models (XML, JSON) to various graph formats (RDF, property graphs) or text. Most real-world datasets contain a certain amount of null values, denoting missing, unknown, or inapplicable information. While some data models allow representing nulls by special tokens, so-called disguised missing values (DMVs, in short) are also frequently encountered: these are values that are not syntactically speaking nulls, but which do, nevertheless, denote the absence, unavailability, or inapplicability of the information. In this work, we tackle the detection of a particular kind of DMV: texts freely entered by human users. This problem is not tackled by DMV detection methods focused on numeric or categoric data; further, it also escapes DMV detection methods based on value frequency, since such free texts are often different from each other, thus most DMVs are unique. We encountered this problem within the ConnectionLens [6,7,8,12] project where heterogeneous data is integrated into large graphs. We present two DMV detection methods for our specific problem: (i) leveraging Information Extraction, already applied in ConnectionLens graphs; and (ii) through text embeddings and classification. We detail their performanceprecision trade-offs on real-world datasets.

International audience; Investigative Journalism (IJ, in short) requires combining highly heterogeneous digital datasets coming from a wide variety of sources. We have developed ConnectionLens, a system that integrates such sources into a single heterogeneous graph and enables users to query the graph using keywords. The first iteration of the system [7] followed a mediator architecture which severely constrained its query scalability. Thus, we fully re-engineered the system, moving it to a warehouse architecture, and replacing its core components (information extraction, data querying, and interactive interfaces), which allowed us to handle uses cases orders of magnitude larger than the previous platform. In a consortium of computer scientists and investigative journalists, we propose to demonstrate ConnectionLens' capability to integrate arbitrary heterogeneous datasets and query them flexibly by means of keywords. Among several scenarios, our main focus will be on a real-world journalistic use case about situations which may lead to Conflicts of Interest between biomedical experts and various organizations, such as corporations, lobbies, etc. The demonstration will showcase the end-to-end data analysis pipeline, illustrate each system component, and the different parameters governing graph creation and querying.

International audience; The era of Big Data and data sharing has lead to very large volumes of data becoming available to users across the world. This data is heterogeneous in its modelling, format and quality. Taking full advantage of such data raises many challenges, in particular related to the integration and the understanding of such data. My PhD thesis, started in January 2021, seeks to develop novel methods to help users without advanced IT skills discover a new dataset, by () building an abstract understanding of the data, as consisting of records and collections, () interpreting or classifying the data based on users' interests, and leveraging Information Extraction and Natural Language tools.

Large (Linked) Open Data are increasingly shared as RDF graphs today. However, such data does not yet reach its full potential in terms of sharing and reuse. We provide new methods to meaningfully summarize data graphs, with a particular focus on RDF graphs. One class of tools for this task are structural RDF graph summaries, which allow users to grasp the different connections between RDF graph nodes. To this end, we introduce our novel RDFQuotient tool that finds compact yet informative RDF graph summaries that can serve as first-sight visualizations of an RDF graph’s structure. We also consider the problem of automatically identifying the k most interesting aggregate queries that can be evaluated on an RDF graph, given an integer k and a user-specified interestingness function. Aggregate queries are routinely used to learn insights from relational data warehouses, and some prior research has addressed the problem of automatically recommending interesting aggregate queries.; Les données ouvertes sont souvent partagées sous la forme de graphes RDF, qui sont une incarnation du principe Linked Open Data (données ouvertes liées). De telles données n’ont toutefois pas atteint leur entier potentiel d’utilisation et de partage. L’obstacle pour ce faire réside principalement au niveau de la capacité des utilisateurs à explorer, découvrir et saisir le contenu et des graphes RDF; cette tâche est complexe car les graphes sont naturellement hétérogènes, et peuvent être à la fois volumineux et complexes. Nous proposons de nouvelles méthodes pour résumer de grands graphes de données, avec un accent particulier sur les graphes RDF. A cette fin, nous avons proposé une nouvelle approché pour la construction de résumés structurels de graphes RDF, à savoir RDFQuotient.Nous considérons aussi le problème d’identifier automatiquement les requêtes d’agrégation les plus intéressantes qui peuvent être évaluées sur un graphe RDF.

International audience; During this internship, we worked on improving an open domain question answering system. We addressed the document selection part which is structured as a text ranking task. The first step was to explore classical methods, also known as sparse retrievers, and to test those algorithms on our evaluation dataset. These methods produced only minor differences in performance. The next step was to employ deep language models, namely the BERT based architectures. A variety of techniques and designs were considered. First, we tackled the lack of data to train such models on the French language, followed by the definition of the problem (classification or ranking), and finally, we addressed the problem of limited text length in BERT-based models. The final results show a 12% improvement in performance over the original model.

In this paper, we consider existential rules, an expressive formalism well adapted to the representation of ontological knowledge, as well as data-to-ontology mappings in the context of ontology-based data integration. The chase is a fundamental tool to do reasoning with existential rules as it computes all the facts entailed by the rules from a database instance. We introduce parallelisable sets of existential rules, for which the chase can be computed in a single breadth-first step from any instance. The question we investigate is the characterization of such rule sets. We show that parallelisable rule sets are exactly those rule sets both bounded for the chase and belonging to a novel class of rules, called pieceful. The pieceful class includes in particular frontier-guarded existential rules and (plain) datalog. We also give another characterization of parallelisable rule sets in terms of rule composition based on rewriting.

International audience; Le BRGM, Service géologique national, est l'établissement public de référence dans les applications des sciences de la Terre pour gérer les ressources et les risques du sol et du sous-sol. Son action est orientée vers la recherche scientifique, l'appui aux politiques publiques et la coopération internationale.Le BRGM développe notamment depuis plusieurs années une expertise reconnue à l’internationale des systèmes d’information dédiés à l’environnement, en particulier sur le domaine de l’eau, du site web portail de référence aux API web de diffusion des données environnementales, en passant par l’indexation de grands volumes de données et la modélisation prédictive. Parmis les outils developés par le BRGM, le portail Hub'eau qui comporte une dizaine d'API Rest chacune spécialisé dans un domaine en relation avec l'eau (qualité des eaux souterraines, rivières, poisson, etc). Pour permettre aux usagers intéressés par les problématiques environnementales de bénéficier des informations que peuvent offrir ces données, nous proposons de concevoir les premières briques d’un service numérique dont la fonction sera de répondre aux questions des utilisateurs, posées en langage naturel, de façon contextualisée et en se basant sur les données réelles du portail Hub'eau.Dans ce rapport nous commençons par une brève description de l'état de l'art en matière de chatbots et de systèmes questions-réponses. Par la suite, nous décrirons l'outil Hub'eau développé par le BRGM, nous aborderons les détails d'une des API du portail en précisant la procédure d'interrogation et le format des données. Nous poursuivrons par la description de la démarche entreprise pour le développement du système décrit ci-dessus. Nous terminerons par la présentation des résultats auxquels nous avons pu aboutir.

International audience; In this report, we will attempt to improve a failure prediction model that is currently used by the French Ministry of Economy and Finance.First, we studied several models and benchmarked them in order to compare our results with those of the articles studied. As a result, we were able to select four models that stood out from the rest, and that we should improve as much as possible.Secondly, we decided to look at the data itself. We realized that our dataset was static. That is, for each row in our table, we had data only for a time T. We therefore decided to add variables, which we will call temporal features, in order to take temporality into account in the model. This addition was more than conclusive, because it allowed us to obtain excellent results that had not been achieved until then.Afterwards, we will proceed with this new dataset. In order to further improve our results, we have started to search by sector of activity. We separatedour dataset into several datasets, the separation being done on the sector of activity of the companies. In doing so, we realized that if we applied different models for each business sector, we would get much better results. Depending on the operational needs, we conclude that this is an area to consider seriously.Finally, we decided to look at the importance of features in our models. To do this, we looked at the importance of the variables in the classifiers, and we realized that only a fraction of the input variables were actually useful, and among those, our temporal features that we added. It would therefore be appropriate to reduce the number of input variables, and to go even further in temporalizing the model on the smallremaining dataset.

International audience; People debate on a variety of topics on online platforms such as Reddit, or Facebook. Debates can be lengthy, with users exchanging a wealth of information and opinions. However, conversations do not always go smoothly, and users sometimes engage in unsound argumentation techniques to prove a claim. These techniques are called fallacies. Fallacies are persuasive arguments that provide insufficient or incorrect evidence to support the claim. In this paper, we study the most frequent fallacies on Reddit, and we present them using the pragma-dialectical theory of argumentation. We construct a new annotated dataset of fallacies, using user comments containing fallacy mentions as noisy labels, and cleaning the data via crowdsourcing. Finally, we study the task of classifying fallacies using neural models. We find that generally the models perform better in the presence of conversational context.We have released the data and the code at github.com/sahaisaumya/informal_ fallacies.

International audience; In this demo, we introduce Exathlon - a new benchmarking platform for explainable anomaly detection over high-dimensional time series. We designed Exathlon to support data scientists and researchers in developing and evaluating learned models and algorithms for detecting anomalous patterns as well as discovering their explanations. This demo will showcase Exathlon's curated anomaly dataset, novel benchmarking methodology, and end-to-end data science pipeline in action via example usage scenarios.

Digital data is a gold mine for modern journalism. However, datasets which interest journalists are extremely heterogeneous, ranging from highly structured (relational databases), semi-structured (JSON, XML, HTML), graphs (e.g., RDF), and text. Journalists (and other classes of users lacking advanced IT expertise, such as most non-governmental-organizations, or small public administrations) need to be able to make sense of such heterogeneous corpora, even if they lack the ability to define and deploy custom extract-transform-load workflows, especially for dynamically varying sets of data sources. We describe a complete approach for integrating dynamic sets of heterogeneous datasets along the lines described above: the challenges we faced to make such graphs useful, allow their integration to scale, and the solutions we proposed for these problems. Our approach is implemented within the ConnectionLens system; we validate it through a set of experiments., 40 pages, 9 figures. arXiv admin note: substantial text overlap with arXiv:2007.12488, arXiv:2009.04283

In this thesis, we propose a human-in-the-loop framework for efficient model development over large data sets. In this framework, we aim to apply active learning algorithms to select a small sequence of data instances for the user to label and derive an accurate model while, at the same time, offering interactive performance in presenting the next data instance for reviewing. However, existing active learning techniques often fail to provide satisfactory performance when built over large data sets. Not only do such models often require hundreds of labeled data instances to reach high accuracy, but retrieving the next instance to label can be time-consuming, making it incompatible with the interactive nature of the human exploration process. To address these issues, we propose the following contributions:1) A new theoretical framework that allows for an efficient implementation of the Generalized Binary Search strategy over kernel classifiers. Compared to previous work, our framework offers both strong theoretical guarantees on performance and efficient implementation in time and space.2) An optimized VS algorithm called OptVS that uses the hit-and-run algorithm for sampling the version space. We also develop a range of sampling optimizations to improve both sample quality and running time. In practice, we observe that OptVS achieves similar or better performance than state-of-the-art version space algorithms while running under 1 second per iteration at all times.3) A new algorithm that leverages the factorization structure provided by the user to create subspacesand factorizes the version space accordingly to perform active learning in the subspaces. We also provide theoretical results on the optimality of our factorized VS algorithm and optimizations for dealing with categorical variables. Our evaluation results show that, for all user patterns considered, our factorized VS algorithm outperforms non-factorized active learners as well as DSM, another factorization-aware algorithm, often by a wide margin while maintaining interactive speed.4) Following the intuitive reasoning behind the user decision-making process, we develop a new human-inspired classification algorithm, called the Factorized Linear Model (FLM), that decomposes the user interest as a combination of low-dimensional convex objects, resulting in an accurate, efficient, and interpretable classifier. In practice, we observe that the FLM classifier achieves comparable or better performance than SVM and another interpretable model, VIPR, over the majority of user interest patterns while taking only a few minutes to train over a large data set of nearly one million points.5) A novel automatically factorized active learning strategy called the Swapping Algorithm. This technique initially employs OptVS to escape the slow convergence of initial iterations and then swaps to an FLM-based strategy to take advantage of its higher classification accuracy. Our evaluation shows that the Swapping Algorithm achieves similar or better performance than non-factorized active learners while approximating the explicitly factorized methods.; Dans cette thèse, nous proposons un cadre "human-in-the-loop" pour le développement efficace de modèles sur de grands ensembles de données. Dans ce cadre, nous visons à appliquer des algorithmes d'apprentissage actif pour sélectionner une petite séquence d'instances de données que l'utilisateur doit étiqueter et dériver un modèle précis, tout en offrant une performance interactive en présentant l'instance de données suivante pour examen. Cependant, les techniques d'apprentissage actif existantes ne parviennent souvent pas à fournir des performances satisfaisantes lorsqu'elles sont construites sur de grands ensembles de données. Non seulement ces modèles nécessitent souvent des centaines d'instances de données étiquetées pour atteindre un niveau de précision élevé, mais la récupération de l'instance suivante à étiqueter peut prendre beaucoup de temps, ce qui la rend incompatible avec la nature interactive du processus d'exploration humain. Pour résoudre ces problèmes, nous proposons les contributions suivantes :1) Un nouveau cadre théorique qui permet une mise en œuvre efficace de la stratégie de recherche binaire généralisée sur les classificateurs à noyau. Par rapport aux travaux précédents, notre cadre offre à la fois de solides garanties théoriques sur les performances et une mise en œuvre efficace en temps et en espace.2) Un algorithme VS optimisé appelé OptVS qui utilise l'algorithme hit-and-run pour échantillonner l'espace des versions. Nous développons également une série d'optimisations de l'échantillonnage pour améliorer à la fois la qualité de l'échantillon et le temps d'exécution. Dans la pratique, nous observons qu'OptVS atteint des performances similaires ou supérieures à celles des algorithmes d'espace de version de l'état de l'art tout en fonctionnant en permanence en dessous d'une seconde par itération.3) Un nouvel algorithme qui exploite la structure de factorisation fournie par l'utilisateur pour créer des sous-espaces et factoriser l'espace de version en conséquence pour effectuer un apprentissage actif dans les sous-espaces. Nous fournissons également des résultats théoriques sur l'optimalité de notre algorithme VS factorisé et des optimisations pour traiter les variables catégorielles. Nos résultats d'évaluation montrent que, pour tous les modèles d'utilisateurs considérés, notre algorithme VS factorisé surpasse les apprenants actifs non factorisés ainsi que DSM, un autre algorithme prenant en compte la factorisation, souvent par une large marge tout en maintenant la vitesse interactive. 4) En suivant le raisonnement intuitif derrière le processus de prise de décision de l'utilisateur, nous développons un nouvel algorithme de classification inspiré par l'homme, appelé le modèle linéaire factorisé (FLM), qui décompose l'intérêt de l'utilisateur comme une combinaison d'objets convexes de faible dimension, résultant en un classificateur précis, efficace et interprétable. En pratique, nous observons que le classificateur FLM atteint des performances comparables ou supérieures à celles du SVM et d'un autre modèle interprétable, VIPR, pour la majorité des modèles d'intérêt de l'utilisateur, tout en prenant seulement quelques minutes pour s'entraîner sur un grand ensemble de données de près d'un million de points. 5) Une nouvelle stratégie d'apprentissage actif factorisé automatiquement appelée l'algorithme de permutation. Cette technique utilise initialement OptVS pour échapper à la convergence lente des itérations initiales, puis passe à une stratégie basée sur FLM pour profiter de sa précision de classification supérieure. Notre évaluation montre que l'algorithme de permutation atteint des performances similaires ou supérieures à celles des apprenants actifs non factorisés tout en s'approchant des méthodes explicitement factorisées.

Faced with an increasing gap between fast growth of data and limited human ability to comprehend data, data analytics tools are now in high demand in many applications across a broad set of domains. In particular, for interactive data exploration systems, an "explore-by-example" framework, which aims to assist the user in performing highly effective data exploration while minimizing the human effort, is becoming increasingly popular. However, the state-of-the-art explore-by-example systems still require a large number of labeled examples to achieve the desired accuracy and cannot handle noisy labels. To address both the slow convergence problem and the label noise problem, in this thesis, we cast the explore-by-example problem in a principled "active learning" framework, and bring the properties of important classes of the user interest to bear on the design of new algorithms and optimizations for active learning-based data exploration. While recent work proposed a polytope-based data space model to filter examples returned by a traditional active learner and to compute an accuracy-based stopping criterion for active learning, our first contribution is to combine the polytope-based model and a traditional active learner into a new Dual-Space Model (\dsm), jointly offering the prediction and sample acquisition functionalities and thereby expediting model convergence. We also provide a set of techniques to improve the interactive performance such that the per-iteration time is maintained within 1 to 2 seconds. Evaluation results using real-world datasets and user interest patterns show that the accuracy of our system is on average 26\% higher than the state-of-the-art explore-by-example systems.Our second contribution is to overcome the slow convergence problem when data exploration is performed in high dimensions. We factorize the high-dimensional space into a set of low-dimensional spaces, build a \dsm\ in each subspace and combine them to be a factorized \dsm\ (\dsmf). We further prove that \dsmf\ achieves a better accuracy-based stopping criterion than \dsm. In case that the user may start exploration with more attributes than those needed in the final model, we propose an online feature selection method that adaptively selects the top-k relevant attributes. Experimental results using real-world workloads show that our system significantly outperforms the state-of-the-art explore-by-example systems in accuracy (19x$\sim$64x accuracy improvement) and convergence speed while maintaining the per-iteration time within 1 to 2 seconds and our online feature selection method improves accuracy from nearly 0 without feature selection, to above 80\%. %We formally define the class of queries that \dsmf\ supports, the decision function it utilizes, and prove that it achieves a better accuracy-based stopping criterion than \dsm.Last but not least, we address the label noise problem when the labels provided by the user are potentially corrupted. We enhance the robustness of our system by integrating advanced data cleansing methods and a refinement of the polytope-based model into \dsmf. The resulting algorithm, referred to as the Robust Dual-Space Model (\rdsmf), is compared to traditional active learning for evaluation since the state-of-the-art explore-by-example systems fail to deal with noisy labels. Experimental results using real-world datasets and user interest patterns show that our proposed algorithm substantially outperforms traditional active learning in accuracy (up to 22x accuracy improvement) while achieving reasonable efficiency for interactive data exploration (1 to 3 seconds per iteration).; Face à un écart grandissant entre la croissance rapide des données et la capacité humaine limitée à comprendre les données, les outils d'analyse de données sont en forte demande dans de nombreuses applications à travers un large éventail de domaines. En particulier, pour les systèmes interactifs d'exploration de données, un cadre « explorer par l’exemple », qui vise à aider un utilisateur humain à effectuer une exploration de données très efficace tout en minimisant son effort, devient de plus en plus populaire. Cependant, ces systèmes de pointe nécessitent toujours un grand nombre d'exemples étiquetés pour obtenir la précision souhaitée et ne peuvent pas gérer les étiquettes bruyantes. Pour résoudre à la fois le problème de la convergence lente et le problème du bruit des étiquettes, dans cette thèse, nous plaçons le problème d’« explorer par l’exemple » dans un cadre d’apprentissage actif fondé sur des principes et apportons les caractéristiques des classes importantes de l'intérêt de l'utilisateur, pour contribuer à la conception de nouveaux algorithmes et à l'optimisation pour l'exploration de données basée sur l'apprentissage actif.Alors que des travaux récents ont proposé un modèle d'espace de données basé sur les polytopes pour filtrer les exemples renvoyés par un apprenant actif traditionnel et pour calculer un critère d'arrêt basé sur la précision pour l'apprentissage actif, notre première contribution vise à combiner le modèle basé sur les polytopes et l’apprenant actif traditionnel en un nouveau modèle à double espace (\dsm), qui offre conjointement les fonctionnalités de prédiction et d'acquisition d'échantillons et ainsi l’accélération de la convergence des modèles. D'ailleurs, nous introduisons un ensemble de techniques pour améliorer les performances interactives de telle sorte que le temps par itération est maintenu entre 1 à 2 secondes. Les résultats de l'évaluation utilisant des jeux de données du monde réel et des patterns d'intérêt des utilisateurs montrent que la précision de notre système est en moyenne 26\% supérieure à celle des systèmes d’explorer par l’exemple de pointe.Notre deuxième contribution consiste à surmonter le problème de la convergence lente lorsque l'exploration des données est effectuée en haute dimension. Nous factorisons l'espace de grande dimension en un ensemble d'espaces de faibles dimensions, construisons un \dsm\ dans chaque sous-espace et les combinons pour former un \dsm\ factorisé (\dsmf). Nous prouvons en outre que \dsmf\ atteint un meilleur critère d'arrêt basé sur la précision que \dsm. Dans le cas où l'utilisateur veut commencer l'exploration avec plus d'attributs que ceux requis dans le modèle final, nous proposons une méthode de sélection de caractéristique en ligne qui donne de manière adaptative les k attributs les plus pertinents. Les résultats expérimentaux montrent que notre système surpasse les systèmes de pointe en termes de précision (entre x19 et x64) et de vitesse de convergence tout en maintenant le temps par itération entre 1 à 2 secondes et notre méthode de sélection de caractéristique en ligne améliore la précision de près de 0 (sans sélection) jusqu’à plus de 80\%.Enfin, nous abordons le problème du bruit lorsque les étiquettes fournies par l'utilisateur sont potentiellement erronées. Nous renforçons la robustesse de notre système en intégrant des méthodes avancées de nettoyage de données et un raffinement du modèle basé sur les polytopes dans \dsmf. L'algorithme obtenu, appelé le modèle à double espace robuste (\rdsmf), est comparé à l'apprentissage actif traditionnel à des fins d'évaluation, car les systèmes d’explorer par l’exemple de pointe sont incapables de gérer les étiquettes bruyantes. Les résultats expérimentaux montrent que notre algorithme améliore considérablement l'apprentissage actif traditionnel en termes de précision (jusqu'à x22) tout en atteignant une efficacité raisonnable pour l'exploration interactive des données (1 à 3 secondes par itération).

Faced with an increasing gap between fast growth of data and limited human ability to comprehend data, data analytics tools are now in high demand in many applications across a broad set of domains. In particular, for interactive data exploration systems, an "explore-by-example" framework, which aims to assist the user in performing highly effective data exploration while minimizing the human effort, is becoming increasingly popular. However, the state-of-the-art explore-by-example systems still require a large number of labeled examples to achieve the desired accuracy and cannot handle noisy labels. To address both the slow convergence problem and the label noise problem, in this thesis, we cast the explore-by-example problem in a principled "active learning" framework, and bring the properties of important classes of the user interest to bear on the design of new algorithms and optimizations for active learning-based data exploration. While recent work proposed a polytope-based data space model to filter examples returned by a traditional active learner and to compute an accuracy-based stopping criterion for active learning, our first contribution is to combine the polytope-based model and a traditional active learner into a new Dual-Space Model (\dsm), jointly offering the prediction and sample acquisition functionalities and thereby expediting model convergence. We also provide a set of techniques to improve the interactive performance such that the per-iteration time is maintained within 1 to 2 seconds. Evaluation results using real-world datasets and user interest patterns show that the accuracy of our system is on average 26\% higher than the state-of-the-art explore-by-example systems.Our second contribution is to overcome the slow convergence problem when data exploration is performed in high dimensions. We factorize the high-dimensional space into a set of low-dimensional spaces, build a \dsm\ in each subspace and combine them to be a factorized \dsm\ (\dsmf). We further prove that \dsmf\ achieves a better accuracy-based stopping criterion than \dsm. In case that the user may start exploration with more attributes than those needed in the final model, we propose an online feature selection method that adaptively selects the top-k relevant attributes. Experimental results using real-world workloads show that our system significantly outperforms the state-of-the-art explore-by-example systems in accuracy (19x$\sim$64x accuracy improvement) and convergence speed while maintaining the per-iteration time within 1 to 2 seconds and our online feature selection method improves accuracy from nearly 0 without feature selection, to above 80\%. %We formally define the class of queries that \dsmf\ supports, the decision function it utilizes, and prove that it achieves a better accuracy-based stopping criterion than \dsm.Last but not least, we address the label noise problem when the labels provided by the user are potentially corrupted. We enhance the robustness of our system by integrating advanced data cleansing methods and a refinement of the polytope-based model into \dsmf. The resulting algorithm, referred to as the Robust Dual-Space Model (\rdsmf), is compared to traditional active learning for evaluation since the state-of-the-art explore-by-example systems fail to deal with noisy labels. Experimental results using real-world datasets and user interest patterns show that our proposed algorithm substantially outperforms traditional active learning in accuracy (up to 22x accuracy improvement) while achieving reasonable efficiency for interactive data exploration (1 to 3 seconds per iteration).; Face à un écart grandissant entre la croissance rapide des données et la capacité humaine limitée à comprendre les données, les outils d'analyse de données sont en forte demande dans de nombreuses applications à travers un large éventail de domaines. En particulier, pour les systèmes interactifs d'exploration de données, un cadre « explorer par l’exemple », qui vise à aider un utilisateur humain à effectuer une exploration de données très efficace tout en minimisant son effort, devient de plus en plus populaire. Cependant, ces systèmes de pointe nécessitent toujours un grand nombre d'exemples étiquetés pour obtenir la précision souhaitée et ne peuvent pas gérer les étiquettes bruyantes. Pour résoudre à la fois le problème de la convergence lente et le problème du bruit des étiquettes, dans cette thèse, nous plaçons le problème d’« explorer par l’exemple » dans un cadre d’apprentissage actif fondé sur des principes et apportons les caractéristiques des classes importantes de l'intérêt de l'utilisateur, pour contribuer à la conception de nouveaux algorithmes et à l'optimisation pour l'exploration de données basée sur l'apprentissage actif.Alors que des travaux récents ont proposé un modèle d'espace de données basé sur les polytopes pour filtrer les exemples renvoyés par un apprenant actif traditionnel et pour calculer un critère d'arrêt basé sur la précision pour l'apprentissage actif, notre première contribution vise à combiner le modèle basé sur les polytopes et l’apprenant actif traditionnel en un nouveau modèle à double espace (\dsm), qui offre conjointement les fonctionnalités de prédiction et d'acquisition d'échantillons et ainsi l’accélération de la convergence des modèles. D'ailleurs, nous introduisons un ensemble de techniques pour améliorer les performances interactives de telle sorte que le temps par itération est maintenu entre 1 à 2 secondes. Les résultats de l'évaluation utilisant des jeux de données du monde réel et des patterns d'intérêt des utilisateurs montrent que la précision de notre système est en moyenne 26\% supérieure à celle des systèmes d’explorer par l’exemple de pointe.Notre deuxième contribution consiste à surmonter le problème de la convergence lente lorsque l'exploration des données est effectuée en haute dimension. Nous factorisons l'espace de grande dimension en un ensemble d'espaces de faibles dimensions, construisons un \dsm\ dans chaque sous-espace et les combinons pour former un \dsm\ factorisé (\dsmf). Nous prouvons en outre que \dsmf\ atteint un meilleur critère d'arrêt basé sur la précision que \dsm. Dans le cas où l'utilisateur veut commencer l'exploration avec plus d'attributs que ceux requis dans le modèle final, nous proposons une méthode de sélection de caractéristique en ligne qui donne de manière adaptative les k attributs les plus pertinents. Les résultats expérimentaux montrent que notre système surpasse les systèmes de pointe en termes de précision (entre x19 et x64) et de vitesse de convergence tout en maintenant le temps par itération entre 1 à 2 secondes et notre méthode de sélection de caractéristique en ligne améliore la précision de près de 0 (sans sélection) jusqu’à plus de 80\%.Enfin, nous abordons le problème du bruit lorsque les étiquettes fournies par l'utilisateur sont potentiellement erronées. Nous renforçons la robustesse de notre système en intégrant des méthodes avancées de nettoyage de données et un raffinement du modèle basé sur les polytopes dans \dsmf. L'algorithme obtenu, appelé le modèle à double espace robuste (\rdsmf), est comparé à l'apprentissage actif traditionnel à des fins d'évaluation, car les systèmes d’explorer par l’exemple de pointe sont incapables de gérer les étiquettes bruyantes. Les résultats expérimentaux montrent que notre algorithme améliore considérablement l'apprentissage actif traditionnel en termes de précision (jusqu'à x22) tout en atteignant une efficacité raisonnable pour l'exploration interactive des données (1 à 3 secondes par itération).

International audience; As enterprise information systems are collecting event streams from various sources, the ability of a system to automatically detect anomalous events and further provide human-readable explanations is of paramount importance. In this paper, we present an approach to integrated anomaly detection (AD) and explanation discovery (ED), which is able to leverage state-of-the-art Deep Learning (DL) techniques for anomaly detection, while being able to recover human-readable explanations for detected anomalies. At the core of the framework is a new human-interpretable dimensionality reduction (HIDR) method that not only reduces the dimensionality of the data, but also maintains a meaningful mapping from the original features to the transformed low-dimensional features. Such transformed features can be fed into any DL technique designed for anomaly detection, and the feature mapping will be used to recover human-readable explanations through a suite of new feature selection and explanation discovery methods. Evaluation using a recent explainable anomaly detection benchmark demonstrates the efficiency and effectiveness of HIDR for AD, and the result that while all three recent ED techniques failed to generate quality explanations on high-dimensional data, our HIDR-based ED framework can enable them to generate explanations with dramatic improvements in the quality of explanations and computational efficiency.

As large Open Data are increasingly shared as RDF graphs today, there is a growing demand to help users discover the most interesting facets of a graph, which are often hard to grasp without automatic tools. We consider the problem of automatically identifying the k most interesting aggregate queries that can be evaluated on an RDF graph, given an integer k and a user-specified interestingness function. Our problem departs from analytics in relational data warehouses in that (i) in an RDF graph we are not given but we must identify the facts, dimensions, and measures of candidate aggregates; (ii) the classical approach to efficiently evaluating multiple aggregates breaks in the face of multi-valued dimensions in RDF data. In this work, we propose an extensible end-to-end framework that enables the identification and evaluation of interesting aggregates based on a new RDF-compatible one-pass algorithm for efficiently evaluating a lattice of aggregates and a novel early-stop technique (with probabilistic guarantees) that can prune uninteresting aggregates. Experiments using both real and synthetic graphs demonstrate the ability of our framework to find interesting aggregates in a large search space, the efficiency of our algorithms (with up to 2.9x speedup over a similar pipeline based on existing algorithms), and scalability as the data size and complexity grow., Comment: Accepted for publication in proceedings of the 2021 International Conference on Management of Data (SIGMOD '21), June 20--25, 2021, Virtual Event, China

International audience; Hybrid complex analytics workloads typically include (i) data management tasks (joins, filters, etc.), easily expressed using relational algebra (RA)-based languages, and (ii) complex analytics tasks (regressions, matrix decompositions, etc.), mostly expressed in linear algebra (LA) expressions. Such workloads are common in a number of areas, including scientific computing, web analytics, business recommendation, natural language processing, speech recognition. Existing solutions for evaluating hybrid complex analytics queriesranging from LA-oriented systems, to relational systems (extended to handle LA operations), to hybrid systems-fail to provide a unified optimization framework for such a hybrid setting. These systems either optimize data management and complex analytics tasks separately, or exploit RA properties only while leaving LA-specific optimization opportunities unexplored. Finally, they are not able to exploit precomputed (materialized) results to avoid computing again (part of) a given mixed (LA and RA) computation. We describe HADAD, an extensible lightweight approach for optimizing hybrid complex analytics queries, based on a common abstraction that facilitates unified reasoning: a relational model endowed with integrity constraints, which can be used to express the properties of the two computation formalisms. Our approach enables full exploration of LA properties and rewrites, as well as semantic query optimization. Importantly, our approach does not require modifying the internals of the existing systems. Our experimental evaluation shows significant performance gains on diverse workloads, from LA-centered ones to hybrid ones.

International audience; Data analytics in the cloud has become an integral part of enterprise businesses. Big data analytics systems, however, still lack the ability to take user performance goals and budgetary constraints for a task, collectively referred to as task objectives, and automatically configure an analytic job to achieve these objectives. This paper presents a data analytics optimizer that can automatically determine a cluster configuration with a suitable number of cores as well as other system parameters that best meet the task objectives. At a core of our work is a principled multi-objective optimization (MOO) approach that computes a Pareto optimal set of job configurations to reveal tradeoffs between different user objectives, recommends a new job configuration that best explores such tradeoffs, and employs novel optimizations to enable such recommendations within a few seconds. We present efficient incremental algorithms based on the notion of a Progressive Frontier for realizing our MOO approach and implement them into a Spark-based prototype. Detailed experiments using benchmark workloads show that our MOO techniques provide a 2-50x speedup over existing MOO methods, while offering good coverage of the Pareto frontier. When compared to Ottertune, a state-of-the-art performance tuning system, our approach recommends configurations that yield 26%-49% reduction of running time of the TPCx-BB benchmark while adapting to different application preferences on multiple objectives.

Cloud data analytics has become an integral part of enterprise business operations for data-driven insight discovery. Performance modeling of cloud data analytics is crucial for performance tuning and other critical operations in the cloud. Traditional modeling techniques fail to adapt to the high degree of diversity in workloads and system behaviors in this domain. In this thesis, we bring recent Deep Learning techniques to bear on the process of automated performance modeling of cloud data analytics, with a focus on Spark data analytics as representative workloads. At the core of our work is the notion of learning workload embeddings (with a set of desired properties) to represent fundamental computational characteristics of different jobs, which enable performance prediction when used together with job configurations that control resource allocation and other system knobs. Our work provides an in-depth study of different modeling choices that suit our requirements. Results of extensive experiments reveal the strengths and limitations of different modeling methods, as well as superior performance of our best performing method over a state-of-the-art modeling tool for cloud analytics; L'analyse des données en utilisant des ressources cloud est désormais omniprésente dans l'activité des entreprises qui s'engagent dans une transformation digitale pour mieux comprendre les données volumineuses dont elles disposent. La modélisation des performances des plateformes cloud utilisées dans ce contexte est une nécessité pour pouvoir garantir une bonne performance des requettes réparties (appelées jobs) ainsi qu'une meilleure gestion des ressources cloud. Les techniques de modélisation traditionnelles ne s'adaptent ni à la diversité de ces jobs ni aux différents comportements des systèmes distribués. Dans cette thèse, nous proposons des techniques récentes de Deep Learning pour pouvoir automatiser cette tâche de modélisation avec un focus en particulier sur la plateforme Spark utilisée pour les calculs distribués. Au coeur de notre travaux de recherche, on présente la notion d'apprentissage d'embeddings, vecteurs capables de décrire de façon compacte les caractéristiques fondamentales des différents jobs. Nous montrerons dans cette thèse comment ces embeddings permettent une meilleure prédiction des performances des jobs sous différentes configurations du système de calculs répartis. Nous aborderons aussi une étude de différents choix de modélisation à base de réseaux de neurones répondant à nos besoins. Les résultats de nos expériences révèlent les forces et les limites des différents choix de modélisation. Nos expériences dévoilent aussi des performances supérieures d'une méthode qu'on propose par rapport à l'état de l'art dans la modélisation des systèmes de gestion de base de données.

International audience; Investigative Journalism (IJ, in short) is staple of modern, democratic societies. IJ often necessitates working with large, dynamic sets of heterogeneous, schema-less data sources, which can be structured, semi-structured, or textual, limiting the applicability of classical data integration approaches. In prior work, we have developed ConnectionLens, a system capable of integrating such sources into a single heterogeneous graph, leveraging Information Extraction (IE) techniques; users can then query the graph by means of keywords, and explore query results and their neighborhood using an interactive GUI. Our keyword search problem is complicated by the graph heterogeneity, and by the lack of a result score function that would enable pruning of the search space. In this work, we describe an actual IJ application studying conflicts of interest in the biomedical domain, and we show how ConnectionLens supports it. Then, we present novel techniques addressing the scalability challenges raised by this application: one allows us to reduce the significant IE costs while building the graph, while the other is a novel, parallel, in-memory keyword search engine, which achieves orders of magnitude speed-up over our previous engine. Our experimental study on the realworld IJ application data confirms the benefits of our contributions.

Cloud data analytics has become an integral part of enterprise business operations for data-driven insight discovery. Performance modeling of cloud data analytics is crucial for performance tuning and other critical operations in the cloud. Traditional modeling techniques fail to adapt to the high degree of diversity in workloads and system behaviors in this domain. In this paper, we bring recent Deep Learning techniques to bear on the process of automated performance modeling of cloud data analytics, with a focus on Spark data analytics as representative workloads. At the core of our work is the notion of learning workload embeddings (with a set of desired properties) to represent fundamental computational characteristics of different jobs, which enable performance prediction when used together with job configurations that control resource allocation and other system knobs. Our work provides an in-depth study of different modeling choices that suit our requirements. Results of extensive experiments reveal the strengths and limitations of different modeling methods, as well as superior performance of our best performing method over a state-of-the-art modeling tool for cloud analytics.

International audience; Digital data sharing leads to unprecedented opportunities to develop data-driven systems for supporting economic activities, the social and political life, and science. Many open-access datasets are RDF graphs, but others are CSV files, Neo4J property graphs, JSON or XML documents, etc. Potential users need to understand a dataset in order to decide if it is useful for their goal. While some datasets come with a schema and/or documentation, this is not always the case. Data summarization or schema inference tools have been proposed, specializing in XML, or JSON, or the RDF data models. In this work, we present a dataset abstraction approach, which () applies on relational, CSV, XML, JSON, RDF or Property Graph data; () computes an abstraction meant for humans (as opposed to a schema meant for a parser); () integrates Information Extraction data profiling, to also classify dataset content among a set of categories of interest to the user. Our abstractions are conceptually close to an Entity-Relationship diagram, if one allows nested and possibly heterogeneous structure within entities.

When active learning (AL) is applied to help the user develop a model on a large dataset through interactively presenting data instances for labeling, existing AL techniques can suffer from two main drawbacks: first, they may require hundreds of labeled data instances in order to reach high accuracy; second, retrieving the next instance to label can be time consuming, making it incompatible with the interactive nature of the human exploration process. To address these issues, we introduce a novel version space based AL algorithm for kernel classifiers, which not only has strong theoretical guarantees on performance, but also allows for an efficient implementation in time and space. In addition, by leveraging additional insights obtained in the user labeling process, we are able to factorize the version space to perform active learning in a set of subspaces, which further reduces the user labeling effort. Evaluation results show that our algorithms significantly outperform state-of-theart version space algorithms, as well as a recent factorization-aware algorithm, for model development over large data sets.

International audience; The performance of query answering in an RDF database strongly depends on the data layout, that is, the way data is split in persistent data structures. We consider answering Basic Graph Pattern Queries (BGPQs), and in particular those with variables (also) in class and property positions, in the presence of RDFS ontologies, both through data saturation and query reformulation. We show that such demanding queries often lead to inefficient query answering on two popular storage layouts, so-called T and CP. We present novel query answering algorithms on the TCP layout, which combines T and CP. In exchange to occupying more storage space, e.g. on an inexpensive disk, TCP avoids the bad or even catastrophic performance that T and/or CP sometimes exhibit. We also introduce summary-based pruning, a novel technique based on existing RDF quotient summaries, which improves query answering performance on the T, CP and the more robust TCP layouts.

International audience; Among existing sources of Open Data, statistical databases published by national and international organizations such as the International Monetary Fund, the United Nations, OECD etc. stand out for their high quality and valuable insights. However, technical means to interact easily with such sources are currently lacking. This article presents an effort to build an interactive Question Answering system for accessing statistical databases structured according to the SDMX (Statistical Data and Metadata Exchange) standard promoted by the abovementioned institutions. We describe the system architectures, its main technical choices, and present a preliminary evaluation. The system is available online.

Data journalism is the field of investigative journalism which focuses on digital data by treating them as first-class citizens. Following the trends in human activity, which leaves strong digital traces, data journalism becomes increasingly important. However, as the number and the diversity of data sources increase, heterogeneous data models with different structure, or even no structure at all, need to be considered in query answering. Inspired by our collaboration with Le Monde, a leading French newspaper, we designed a novel query algorithm for exploiting such heterogeneous corpora through keyword search. We model our underlying data as graphs and, given a set of search terms, our algorithm nds links between them within and across the heterogeneous datasets included in the graph. We draw inspiration from prior work on keyword search in structured and unstructured data, which we extend with the data heterogeneity dimension, which makes the keyword search problem computationally harder. We implement our algorithm and we evaluate its performance using synthetic and real-world datasets., Informal publication only

International audience; Data integration is a broad area of data management research. It has lead to the development of many useful tools and concepts, each appropriate in a certain class of applicative settings.We consider the setting in which data sources have heterogeneous data models. This setting is of increasing relevance, as the (once predominant) relational databases are supplemented by data exchanged in formats such as JSON or XML, graphs such as Linked Open (RDF) data, or matrix (numerical) etc. We describe two lines of work in this setting. The first aims on improving performance in a polystore setting, where data sources are queried through a structure, composite query language; the focus here is on dramatically improving performance through the use of view-based rewriting techniques. The second data integration setting assumes that sources are much too heterogeneous for structured querying and thus, explore keyword-based search in an integrated graph built from all the available data.Designing and setting up data integration architectures remains a rather complex task; data heterogeneity makes it all the more challenging. We believe much remains to be done to consolidate and advance in this area in the future.

Big data applications increasingly involve diverse datasets, conforming to different data models. Such datasets are routinely hosted in heterogeneous stores, each capable of handling one or a few data models, and each efficient for some, but not all, kinds of data processing. Systems capable of exploiting disparate data in this fashion are usually termed polystores. A current limitation of polystores is that applications are written taking into account which part of the data is stored in which store and how. This fails to take advantage of ( i ) possible redundancy, when the same data may be accessible (with different performance) from distinct data stores; ( ii ) previous query results (in the style of materialized views), which may be available in the stores. We propose to demonstrate ESTOCADA [4], a novel approach that can be used in a polystore setting to transparently enable each query to benefit from the best combination of stored data and available processing capabilities. The system leverages recent advances in the area of view-based query rewriting under constraints, which we use to describe the various data models and stored data.

International audience