Keywords
Computer Science and Digital Science
- A2.1. Programming Languages
- A2.1.1. Semantics of programming languages
- A2.1.4. Functional programming
- A2.1.6. Concurrent programming
- A2.4. Formal method for verification, reliability, certification
- A2.4.1. Analysis
- A2.4.2. Model-checking
- A2.4.3. Proofs
- A3.1. Data
- A3.1.1. Modeling, representation
- A3.1.2. Data management, quering and storage
- A3.1.3. Distributed data
- A3.1.4. Uncertain data
- A3.1.5. Control access, privacy
- A3.1.6. Query optimization
- A3.1.7. Open data
- A3.1.8. Big data (production, storage, transfer)
- A3.1.9. Database
- A3.2.1. Knowledge bases
- A3.2.2. Knowledge extraction, cleaning
- A3.2.3. Inference
- A3.2.4. Semantic Web
- A4.7. Access control
- A4.8. Privacy-enhancing technologies
- A7. Theory of computation
- A7.2. Logic in Computer Science
- A9.1. Knowledge
- A9.2. Machine learning
- A9.7. AI algorithmics
- A9.8. Reasoning
Other Research Topics and Application Domains
- B6.1. Software industry
- B6.3.1. Web
- B6.3.4. Social Networks
- B6.5. Information systems
- B9.5.1. Computer science
- B9.5.6. Data science
- B9.10. Privacy
1 Team members, visitors, external collaborators
Research Scientists
- Joachim Niehren [Team leader, Inria, Senior Researcher, HDR]
- Mikael Monet [Inria, Researcher, from Oct 2020]
Faculty Members
- Iovka Boneva [Université de Lille, Associate Professor]
- Florent Capelli [Université de Lille, Associate Professor]
- Aurélien Lemay [Université de Lille, Associate Professor, HDR]
- Charles Paperman [Université de Lille, Associate Professor]
- Sylvain Salvati [Université de Lille, Professor, HDR]
- Slawomir Staworko [Université de Lille, Associate Professor, HDR]
- Sophie Tison [Université de Lille, Professor, HDR]
PhD Students
- Antonio Al Serhali [Inria, from Oct 2020]
- Nicolas Crosetti [Inria]
- Paul Gallot [Inria]
- Jose Martin Lozano [Université de Lille]
- Momar Ndiouga Sakho [Université de Lille, until Aug 2020]
- Claire Soyez-Martin [Inria, from Sep 2020]
Technical Staff
- Antonio Al Serhali [Inria, Engineer, until Sep 2020]
- Chérif Amadou Ba [Inria, Engineer, from Sep 2020]
- Momar Ndiouga Sakho [Inria, Engineer, from Sep 2020]
Interns and Apprentices
- Corentin Barloy [École Normale Supérieure de Paris, from Aug 2020 until Sep 2020]
- Leo Beuque [École Centrale de Lille, until Feb 2020]
- Aymeric Come [Inria, from Jul 2020 until Sep 2020]
- Amine Laabi [Université de Lille, from Jul 2020 until Aug 2020]
- Claire Soyez-Martin [Université de Lille, from Feb 2020 until Aug 2020]
Administrative Assistant
- Nathalie Bonte [Inria]
Visiting Scientist
- Corentin Barloy [École Normale Supérieure de Paris, from Oct 2020]
2 Overall objectives
We will develop algorithms for answering logical querying on heterogeneous linked data collections in hybrid formats, distributed programming languages for managing dynamic linked data collections and workflows based on queries and mappings, and symbolic machine learning algorithms that can link datasets by inferring appropriate queries and mappings.
2.1 Presentation
The following three items summarize our main research objectives.
- Querying Heterogeneous Linked Data We develop new kinds of schema mappings for semi-structured datasets in hybrid formats including graph databases, RDF collections, and relational databases. These induce recursive queries on linked data collections for which we will investigate evaluation algorithms, containment problems, and concrete applications.
- Managing Dynamic Linked Data In order to manage dynamic linked data collections and workflows, we will develop distributed data-centric programming languages with streams and parallelism, based on novel algorithms for incremental query answering, study the propagation of updates of dynamic data through schema mappings, and investigate static analysis methods for linked data workflows.
- Linking Data Graphs Finally, we will develop symbolic machine learning algorithms, for inferring queries and mappings between linked data collections in various graphs formats from annotated examples.
3 Research program
3.1 Background
The main objective of Links is to develop methods for querying and managing linked data collections. Even though open linked data is the most prominent example, we will focus on hybrid linked data collections, which are collections of semi-structured datasets in hybrid formats: graph-based, rdf, relational, and NoSQL. The elements of these datasets may be linked, either by pointers or by additional relations between the elements of the different datasets, for instance the “same-as” or “member-of” relations as in rdf.
The advantage of traditional data models is that there exist powerful querying methods and technologies that one might want to preserve. In particular, they come with powerful schemas that constraint the possible manners in which knowledge is represented to a finite number of patterns. The exhaustiveness of these patterns is essential for writing of queries that cover all possible cases. Pattern violations are excluded by schema validation. In contrast, rdf schema languages such as rdfs can only enrich the relations of a dataset by new relations, which also helps for query writing, but which cannot constraint the number of possible patterns, so that they do not come with any reasonable notion of schema validation.
The main weakness of traditional formats, however, is that they do not scale to large data collections as stored on the Web, while the rdf data models scales well to very big collections such as linked open data. Therefore, our objective is to study mixed data collections, some of which may be in rdf format, in which we can lift the advantages of smaller datasets in traditional formats to much larger linked data collections. Such data collections are typically distributed over the internet, where data sources may have rigid query facilities that cannot be easily adapted or extended.
The main assumption that we impose in order to enable the logical approach, is that the given linked data collection must be correct in most dimensions. This means that all datasets are well-formed with respect to their available constraints and schemas, and clean with respect to the data values in most of the components of the relations in the datasets. One of the challenges is to integrate good quality rdf datasets into this setting, another is to clean the incorrect data in those dimensions that are less proper. It remains to be investigated in how far these assumptions can be maintained in realistic applications, and how much they can be weakened otherwise.
For querying linked data collections, the main problems are to resolve the heterogeneity of data formats and schemas, to understand the efficiency and expressiveness of recursive queries, that can follow links repeatedly, to answer queries under constraints, and to optimize query answering algorithms based on static analysis. When linked data is dynamically created, exchanged, or updated, the problems are how to process linked data incrementally, and how to manage linked data collections that change dynamically. In any case (static and dynamic) one needs to find appropriate schema mappings for linking semi-structured datasets. We will study how to automatize parts of this search process by developing symbolic machine learning techniques for linked data collections.
3.2 Research axis: Querying Data Graphs
Linked data is often abstracted as datagraphs: nodes carry information and edges are labeled. Internet, the semantic web, open data, social networks and their connections, information streams such as twitter are examples of such datagraphs. An axis of Links is to propose methods and tools so as to extract information from datagraphs. We dwell in a wide spectrum of tools to construct these methods: circuits, compilation, optimization, logic, automata, machine learning. Our goal is to extend the kinds of information that can be extracted from datagraphs while improving the efficiency of existing ones.
This axis is split within two themes. The first one pertains to the use of lwo level representation by means of circuits to compute efficiently complex numerical aggregates that will find natural applications in AI. The second one proposes to explore path oriented query language and more particularly their efficient evaluation by means of efficient compilation and machine learning methods so as to have manageable statistics.
3.2.1 AI: Circuits for Data Analysis
Circuits are consice representations of data sets that recently found a unifying interest in various areas of artificial intelligence. A circuit may for instance represent the answer set of a database query as a dag whose operators are disjoint unions (for disjunction) and cartesian products (for conjunction). Similarly, it may also represent the set of all matches of a pattern in a graph. The structure of the circuit may give rise to efficient algorithms to process large data sets based on representation that are often much smaller. Among others, such applications range from knowledge representation/compilation, counting the number of solutions of queries, efficient query answering, factorized databases.
In a first line of research, we want to study novel problems on circuits, in which database queries are relevant to data analysis tasks from artificial intelligence, in machine learning or data mining in particular. In particular we propose to study optimiation problems on answer sets of database queries based on circuits, i.e. how to find obtimal solutions in the answer set for a given set of conditions. Decompressing small circuits into large answer sets would make the optimization problem unfeasible in many cases. We believe that circuits can structure certain optimization problems in such a way that it can be phrased concisely and then solved efficiently.
Second, we propose to develop a tighter integration between circuits and databases. Indeed query-related circuits are generally produced from a database. This requires that the data is copied within the circuits. This memory cost is accompanied with the loss of the environment of the DBMS which allows many optimization and uses many low level optimizations that are hard to implement. We propose then to encode circuits directly within the database using materialized views and index structures. We shall also develop the required computational tools for maintaining and exploiting these embedded circuits.
3.2.2 Path Query Optimization
Graph databases are easily queries using path descriptions. Most often these paths are described by means of regular expressions. This makes path queries difficult as the use of Kleene star makes them recursive. In relational DBMS, recursion is almost never used and it is not advised to use it. The natural theoretical tool that pertains to recursion in the context of relational data Datalog. There has been a wealth of optimization algorithms that have been proposed for queries written in Datalog. We propose to use Datalog as a low level language to which we will compile path queries of various kinds. The idea is that the compiler will try to obtain Datalog programs that will have low execution complexity taking advantages of optimization techniques such as magic supplementary set rewriting, pre-computed indexes and also statistics computed from the graph. Our goal is to develop a compiler that will be able to efficiently evaluates path queries on large graphs which in particular will explore only a part of it.
3.3 Research axis: Monitoring Data Graphs
Traditional database applications are programs that interact with database via updates and queries. We are interested in developing programming language techniques so as to interact with datagraphs rather than with traditional relational databases. Moreover, we shall take into account the dynamic aspects of datagraphs which shall evolve through updates. The methods we shall develop will monitor changes in datagraphs and react according to the modifications.
3.3.1 Functional Programming Languges for Data Graphs
The first question is which kind of programming language to use to enable monitoring processes for data graphs based on query answering. While languages of path queries found quite some interest on data graphs, less attention has been given to the programming language tasks, that needed to be solved to produce structured output and to compose various queries with structured output into a pipeline. We believe that transferring the generalization of ideas developed for data trees in the context of XML to data graphs will allow to solve such problems in a systematic manner.
Our approach will consist in developing a functionl programming language based on first principles (the lambda calculus, graph navigation, logical connective) that generalizes full XPath 3.0 to the context of graphs. Here we can rely on own previous work for data trees, such as the language X-Fun and -XP. After the language for data graphs is designed we shall study its behavior empirically by means of an implementation. This implementation will help us to design optimization methods so as to evaluate the queries in that language. We think that in the context of querying functional programs play a central which means that the query language will not allow side effects when computing. This will allow us to use a wealth of techniques so as to optimize the computation. Indeed, we can try to compile data structures to imperative ones when possible and also exploit possibilities of parallel executions in certain cases. Functional programming comes with nice verification techniques that we are going to use in several contexts: (i) in optimizing queries (e.g. stop the evaluation when it is possible to know that no more data can contribute to the output) and (ii) to verify that the query behaves correctly. The verification methods we shall focus on will be mainly related to automata and transducers.
Finally we shall also develop a programming language that allows to describe services that use datagraphs as a backend for storing data. Here again, functional programming seems a good candidate, we would need however to orchestrate the concurrent executions of queries so as to ensure the correct behavior of services. This means that we should have concurrent constructs that are built in the language. The high level of concurrence enabled by the notion of futures seems an interesting candidate to adapt to the context of service orchestration.
3.3.2 Hyperstreaming Program Evaluation
Complex-event processing requires to monitor data graphs that are produced on input streams and to write data graphs to some output stream, which can then be used as inputs again. A major problem here is to reduce the high risk of blocking, which arises when the writing of some of the output stream suspends on a data value that will become available only in the future on some input stream. In such cases, all monitoring processes reading the output stream may have to suspend as well. In order to reduce the risc of blocking, we propose to develop the hyperstreaming approach further, of which we layed the foundations in the evaluation period based on automata techniques. The idea is to generalize streams to hyperstreams, i.e. to add holes to streams that can be filled by some other stream in the future. In order to avoid suspension as possible, a monitoring procesess on hyperstream must then be able to jump over the holes, and to perform some speculative compuation. The objective for the next period are to develop tools for hyperstreaming query answering and to lift these to hyperstreaming program evaluation. Furthermore, on the conceptual side, the notion of certain query answers on hyperstreams needs to be lifted to certain program outputs on hyperstreams.
3.4 Research axis: Graph Data Integration
We intend to continue to develop tools for integration of linked data with RDF being their principal format. Because from its conception the main credo of RDF has been “just publish your data”, the problem at hand faces two important challenges: data quality and data heterogeneity.
3.4.1 Data Quality with Schemas and Repairing with Inference
The data quality of RDF may suffer due to a number of reasons. Impurities may arise due to data value errors (misspellings, errors during data entry etc.). Such data quality problems have been thoroughly investigated in literature for relational databases and solutions include dictionary methods etc. However, it remains to be seen if the challenges of adapting the existing solutions for relational databases can be easily addressed.
One particular challenge comes from the fact that RDF allows a higher degree of structural freedom in how information is represented as opposed to relation databases, where the choice is strongly limited to flat tables. We plan to investigate suitability of existing data cleaning methods to tackle the problems of data value impurities in RDF. The structural freedom of RDF is a source of data quality issues on its own. With the recent emergence of schema formalisms for RDF, it becomes evident that significant parts of existing RDF repositories do not necessarily satisfy schemas prepared by domain experts.
In the first place, we intend to investigate defining suitable measures of quality for RDF documents. Our approaches will be based on a schema language, such as ShEx and SHACL, and we shall explore suitable variants of graph alignment and graph edit distance to capture similarity between the existing RDF document and its possible repaired versions that satisfy the schema.
The central issue here is repairing an RDF document w.r.t. schema by identifying essential fragments of the RDF that fail to satisfy the schema. Once such fragments are identified, repairing actions can be applied however there might be a significant number of alternatives. We intend to explore enumeration approaches where the space of repairing alternatives is intelligently browsed by the user and the most suitable one chosen. Furthermore, we intend to propose a rule language for choosing the most suitable repairing action and will investigate inference methods to derive from interactions with user the optimal order in which various repairing actions are presented to the user and derive the rules for the choice of the preferred repairing action for repeating types of fragments that do not satisfy the schema.
3.4.2 Integration and Graph Mappings with Schemas and Inference
The second problem pertaining to integration of RDF data sources is their heterogeneity. We intend to continue to identify and study suitable classes of mappings between RDF documents conforming to potentially different and complementary schemas. We intend to assist the user in constructing such mappings by developing rich and expressive graphical languages for mappings. Also, we wish to investigate inference of RDF mappings with the active help of an expert user. We will need to define interactive protocols that allows the input to be sufficiently informative to guide the inference process while avoiding the pitfalls of user input being too ambiguous and causing combinatorial explosion. We intend to identify
RDF Data Quality. Approach based on a schema language (ShEx or SHACL) used to identify errors and giving a notion of a measure of quality of an RDF database. Impurities in RDF may come from data value errors (misspellings etc.) but also from the fact that RDF imposes fewer constraints on how data is structured which is a consequence of a significantly different use philosophy (just publish your data anyway you want). Repairing of RDF errors would be modeled with a localized rules (transformations that operate within a small radius of an affected node) and if several rules apply, preferences are used to identify the most desirable one. Both the repairing rules and preferences can be inferred with the help of inference algorithms in an interactive setting. Smart tools for LOD integration. Assuming that the LOD sources are of good quality, we want to build tools that assist the user in constructing mappings that integrate data in the user database. For this, we want to define inference algorithms which are guided by schemas, and which are based on comprehensible interactions with the user. For this, we need to define interactions that are rich enough to inform the algorithm, while simple enough to be understandable by a non-expert user. In particular, that means that we need to present data (nodes in a graph for instance) in a readable way. Also, we want to investigate how the - possibly inferred - schema can be used to guide the inference.
4 Application domains
4.1 Linked data integration
There are many contexts in which integrating linked data is interesting. We advocate here one possible scenario, namely that of integrating business linked data to feed what is called Business Intelligence. The latter consists of a set of theories and methodologies that transform raw data into meaningful and useful information for business purposes (from Wikipedia). In the past decade, most of the enterprise data was proprietary, thus residing within the enterprise repository, along with the knowledge derived from that data. Today's enterprises and businessmen need to face the problem of information explosion, due to the Internet's ability to rapidly convey large amounts of information throughout the world via end-user applications and tools. Although linked data collections exist by bridging the gap between enterprise data and external resources, they are not sufficient to support the various tasks of Business Intelligence. To make a concrete example, concepts in an enterprise repository need to be matched with concepts in Wikipedia and this can be done via pointers or equalities. However, more complex logical statements (i.e. mappings) need to be conceived to map a portion of a local database to a portion of an rdf graph, such as a subgraph in Wikipedia or in a social network, e.g. LinkedIn. Such mappings would then enrich the amount of knowledge shared within the enterprise and let more complex queries be evaluated. As an example, businessmen with the aid of business intelligence tools need to make complex sentimental analysis on the potential clients and for such a reason, such tools must be able to pose complex queries, that exploit the previous logical mappings to guide their analysis. Moreover, the external resources may be rapidly evolving thus leading to revisit the current state of business intelligence within the enterprise.
4.2 Data cleaning
The second example of application of our proposal concerns scientists who want to quickly inspect relevant literature and datasets. In such a case, local knowledge that comes from a local repository of publications belonging to a research institute (e.g. HAL) need to be integrated with other Web-based repositories, such as DBLP, Google Scholar, ResearchGate and even Wikipedia. Indeed, the local repository may be incomplete or contain semantic ambiguities, such as mistaken or missing conference venues, mistaken long names for the publication venues and journals, missing explanation of research keywords, and opaque keywords. We envision a publication management system that exploits both links between database elements, namely pointers to external resources and logical links. The latter can be complex relationships between local portions of data and remote resources, encoded as schema mappings. There are different tasks that such a scenario could entail such as (i) cleaning the errors with links to correct data e.g. via mappings from HAL to DBLP for the publications errors, and via mappings from HAL to Wikipedia for opaque keywords, (ii) thoroughly enrich the list of publications of a given research institute, and (iii) support complex queries on the corrected data combined with logical mappings.
4.3 Real-time complex event processing
Complex event processing serves for monitoring nested word streams in real time. Complex event streams are gaining popularity with social networks such as with Facebook and Twitter, and thus should be supported by distributed databases on the Web. Since this is not yet the case, there remains much space for future industrial transfer related to Links' second axis on dynamic linked data.
5 Highlights of the year
All recently hired permanent members of Links published papers in major conferences on database theory, artificial intelligence, and computer science theory:
- PODS 2021: Principles of Database Systems accepted the paper “Stackless Processing of Streamed Trees” by Corentin Barloy, Filip Murlak and Charles Paperman 3. This is the top-most database theory conference. Corentin is starting his PhD project in Links with Charles.
-
AAAI 2021 Conference on Artificial Intelligence. Two papers accepted.
- Florent Capelli et al. “Certifying Top-Down Decision-DNNF Compilers” 6. Cooperation avec Pierre Marquis de Lens. Furthermore, Florent got an ANR project accepted on related topics.
- Mikaël Monet et al. “The Tractability of SHAP-Score-Based Explanations over Deterministic and Decomposable Boolean Circuits” 2. Furthermore, Mikaël was hired as Inria Junior Researcher this year.
- MFCS 2020: Mathematical Foundations of Computer Science. The paper “Linear high-order deterministic tree transducers with regular look-ahead” by Paul Gallot, Aurélien Lemay, and Sylvain Salvati8 was published by the conference. This contribution will be part of Paul's PhD thesis.
6 New software and platforms
6.1 New software
6.1.1 ShEx validator
- Name: Validation of Shape Expression schemas
- Keywords: Data management, RDF
- Functional Description: Shape Expression schemas is a formalism for defining constraints on RDF graphs. This software allows to check whether a graph satisfies a Shape Expressions schema.
-
Release Contributions:
ShExJava now uses the Commons RDF API and so support RDF4J, Jena, JSON-LD-Java, OWL API and Apache Clerezza. It can parse ShEx schema in the ShEcC, ShEJ, ShExR formats and can serialize a schema in ShExJ.
To validate data against a ShExSchema using ShExJava, you have two different algorithms: - the refine algorithm: compute once and for all the typing for the whole graph - the recursive algorithm: compute only the typing required to answer a validate(node,ShapeLabel) call and forget the results.
-
URL:
http://
shexjava. lille. inria. fr/ - Contact: Iovka Boneva
6.1.2 gMark
- Name: gMark: schema-driven graph and query generation
- Keywords: Semantic Web, Data base
- Functional Description: gMark allow the generation of graph databases and an associated set of query from a schema of the graph.gMark is based on the following principles: - great flexibility in the schema definition - ability to generate big size graphs - ability to generate recursive queries - ability to generate queries with a desired selectivity
-
URL:
https://
github. com/ graphMark/ gmark - Contact: Aurélien Lemay
6.1.3 SmartHal
- Keyword: Bibliography
- Functional Description: SmartHal is a better tool for querying the HAL bibliography database, while is based on Haltool queries. The idea is that a Haltool query returns an XML document that can be queried further. In order to do so, SmartHal provides a new query language. Its queries are conjunctions of Haltool queries (for a list of laboratories or authors) with expressive Boolean queries by which answers of Haltool queries can be refined. These Boolean refinement queries are automatically translated to XQuery and executed by Saxon. A java application for extraction from the command line is available. On top of this, we have build a tool for producing the citation lists for the evaluation report of the LIFL, which can be easily adapter to other Labs.
-
URL:
http://
smarthal. lille. inria. fr/ - Contact: Joachim Niehren
6.1.4 QuiXPath
- Keywords: XML, NoSQL, Data stream
- Scientific Description: The QuiXPath tools supports a very large fragment of XPath 3.0. The QuiXPath library provides a compiler from QuiXPath to FXP, which is a library for querying XML streams with a fragment of temporal logic.
- Functional Description: QuiXPath is a streaming implementation of XPath 3.0. It can query large XML files without loading the entire file in main memory, while selecting nodes as early as possible.
-
URL:
https://
project. inria. fr/ quix-tool-suite/ - Contact: Joachim Niehren
6.1.5 X-FUN
- Keywords: Programming language, Compilers, Functional programming, Transformation, XML
- Functional Description: X-FUN is a core language for implementing various XML, standards in a uniform manner. X-Fun is a higher-order functional programming language for transforming data trees based on node selection queries.
- Contact: Joachim Niehren
- Participants: Joachim Niehren, Pavel Labath
6.1.6 ShapeDesigner
- Name: ShapeDesigner
- Keywords: Validation, Data Exploration, Verification
- Functional Description: ShapeDesigner allows construct a ShEx or SHACL schema for an existing dataset. It combines algorithms to analyse the data and automatically extract shape constraints, and to edit and validate shape schemas.
-
URL:
https://
gitlab. inria. fr/ jdusart/ shexjapp - Contact: Jeremie Dusart
7 New results
7.1 Querying Data Graphs
7.1.1 Circuits for Data Analysis in Artificial Intelligence
Knowledge compilation to Boolean circuits is a general technique in artificial intelligence to obtain tractable algorithms for subclasses of algorithmic problems that are computationally hard. For instance, a variant of Yannakakis' algorithm can be used to compile acyclic conjunctive database queries to Boolean circuits. These will then be decomposable and deterministic, and thus tractable in polynomial time, while for the general class of conjunctive queries, testing the existence of a query answer on a relational database is coNP-complete. Another class of instances, where knowledge complilation is used in AI, concern satisfiablity problems. Beside of satisfiability, knowledge compilation is equally relevant to aggregation and enumeration problems.
In their article in Theory of Computing Systems 11, Capelli, Monet et al. present a systematic picture connecting Boolean circuits to width measures through upper and lower complexity bounds. This is joined work with the Inria Valda team. In particular, their upper bounds show that bounded-treewidth circuits can be constructively converted to special circuits known as d-SDNNFs, in time linear in the circuit size and singly exponential in the treewidth. A much more general survey of complexity question in artificial intelligence is given in a book chapter by Tison et al. 27.
Capelli et al. present at AAAI 6 a method to certify the output knowledge compilers and #SAT-solvers. This is a cooperation with Université de Lens. The idea is to output a certificate that can be checked in polynomial time and can be used to certify that a given CNF formula has K models. Their experiments were encouraging showing that a large majority of CNF formulas for which the #SAT-solver D4 terminates have certificates that can be checked more quickly than the compilation time.
In their article in Discrete Applied Mathematics, Capelli et al. 14 study the problem of faster enumerating models of models of DNF formulas. The aim is to provide enumeration algorithms with a delay that depends polynomially on the size of each model and not on the size of the formula. In particular, they provide a constant delay algorithm for -DNF formulas with fixed .
7.1.2 Uncertainty and Explanations
Monet et al. 19 propose in a paper at NeurIPS a new formalization of the interpretability of classes of models of machine learning algorithms based on computational complexity theory. This work is done in cooperation with the University of Chile. They can prove in their framework that shallow neural networks are more interpretable than deeper neural networks.
Monet et al. 2 study in a paper at AAAI Shapely values for providing explanations to classification results over machine learning models. This work is also done in cooperation with Chile, but now with the Pontifical Catholic University of Chile. While in general computing Shapley values is a computationally intractable problem, it has recently been claimed that the SHAP-score can be computed in polynomial time over the class of decision trees. They show that the SHAP-score can be computed in polynomial time over deterministic and decomposable Boolean circuits.
7.1.3 Path Query Optimization
Niehren, Salvati et al. 25 propose a new algorithm for answering nested regular path queries on data graphs efficiently. Previous jumping algorithms were limited to data trees, while the new jumping evaluator can be applied to data graphs. This generalization is obtained by a novel compilation scheme of path queries to datalog programs.
7.2 Monitoring Data Graphs
7.2.1 Functional Programming Languges for Data Trees
Gallot, Lemay, and Salvati 8 introduced high-order deterministic tree transducers at the 45th International Symposium on Mathematical Foundations of Computer Science (MFCS). This is a natural generalization of known models top-down tree transductions including macro tree transuders and streaming tree transducers. They show that the class of linear high-order tree transducers with look-ahead captures the functional tree-to-tree transformations definable in monadic second-order logic. They also give a specialized procedure for the composition of those transducers that preserves linearity.
Paperman et al. 13 present an article at Logical Methods in Computer Science, in which they study the continuity of functional transducers on words. This is an international cooperation with Chicago and Paris.
7.2.2 Query Answering on Streams
Complex event processing requires to answer queries on streams of complex events, i.e. nested words or, equivalently, linearizations of data trees, but also to produce dynamically evolving data structures as output.
Niehren and Boneva supervised the PhD thesis of Sakho 29 on certain query answering on hyperstreams. They studied the complexity of hyperstreaming query evaluation in a articel published at Information and computation 12. While it is generally in EXP, the complexity goes down to P-time when representing queries by deterministic automata on nested words, and restricting hypestreams to be linear.
In an article published at Algorithms 9 extending on a paper published at CSR 21, they could show that regular path queries on XML documents in the usual XPathMark benchmark can be compiled to reasonably small deterministic automata on nested words. For this they propose new compilers to the novel class of deterministic stepwise hedge automata and proposed a minimization algorithm for them. We note that streaming evaluators for such automata are heavily stack based.
Paperman with his future PhD student Barloy study stackless stream processing for nested words in a cooperation with the University of Warsaw (Poland) 3. They characterize in a paper accepted at the International Conference of Foundations of Database Systems (PODS) the subclass of regular path queries that can be evaluated stacklessly — with and without registers.
7.3 Graph Data Integration
Staworko and Boneva supervised the PhD thesis of Lozano 28 on data exchange from relational database to RDF graphs subject to shape schemas in ShEx. In 26, they show that the consistency problem is coNP-complete, i.e. checking whether every source instance of the relational database admits a target solution, i.e. an RDF graph that satisfies the source-to-target dependencies. They also study the problem of certain query answering, of finding answer of any target solution. For this they introduce the notion of universal simulation solution that allows to compute certain query answers for forwards path queries.
In a cooperation with the Unversity of Oviedo in Spain, Boneva and Staworko conducted a usability experiment on three different graph schema languages for heterogeneous data mapping 15. Their results show that users of our own language ShExML tend to perform better than those of YARRRML and SPARQL-Generate.
7.4 Others
Paperman et al. 13 published a paper on polynomial recursive sequence at the 47th International Colloquium on Automata, Languages and Programming (ICALP). For researching this results, Paperman invited the 4 other authors for a 5 day working meeting in Lille. ICALP is one of the major conferences in theoretical computer science, so this result could be marked as another highlight of the year.
8 Partnerships and cooperations
8.1 International Initiatives
Declared Inria international partners
- Saint Petersburg, Russia Salvati and Niehren cooperate with the University of Saint-Petersburg following a visit of R. Azimov leading to a comon publication at BDA 2020 25. This cooperation was funded by a invitation for R. Azimov by the Cristal lab in 2019.
Informal international partners
- Santiago, Chile Monet cooperates with Marcelo Arenas and Pablo Berceló from the Pontifical Catholic University of Chile and with Luca Bertossi from Adolfo Ibáñez University (also Chile) on counting problems for incomplete databases and on the computation of SHAP-score explainations for circuit classes from knowledge compilation. This yield joint publications at NeuIPS 2020 19 and AAAI 2021 2.
- Warsaw, Poland Paperman cooperates with Filip Murlak on query evaluation on streams. A joint paper is accepted for publication at PODS 2021 3.
- Wroclaw, Poland Staworko has regular exchange with Piotr Wieczorek from the University of Wroclaw, which lead to a joined publication at PODS 2019.
- Tel Aviv, Israel Monet also has regular exchanges with Benny Kimelfeld from Technion (Israel) and Daniel Deutch from Tel Aviv University on computing Shapley values for database query answers.
8.2 International research visitors
8.2.1 Visits of international scientists
- Rustam Azimov Saint-Petersburg State University, Russia, 3 months visit Oct–Dec 2020. Funded by the French-Russian Ambassady. Cancelled due to the Covid pandemic.
- Nofar Cameli Technion, Israel. Links online seminar. Dec 14, 2020.
- Alexandre Vigny Bremen University, Germany. Links online seminar. Dec 10, 2020.
- Pierre Pradic Oxford University, United Kingdom. Links online seminar. Dec 4, 2020.
8.2.2 Sabbatical programme
- Florent Capelli Inria delegation, 2019–2020.
- Slawek Staworko Inria semi-delegation, 2020–2021.
8.3 National initiatives
ANR JCJC KCODA
Participants: Florent Capelli, Charles Paperman, Sylvain Salvati.
- Duration: 2021–2025
- Objectives: The aim of KCODA is to study how succinct representations can be used to efficiently solve modern optimization and AI problems that use a lot of data. We suggest using data structures from the field of compilation of knowledge that can represent large datasets succinctly by factoring certain parts while allowing efficient analysis of the represented data. The first goal of KCODA is to understand how one can efficiently solve optimization and training problems for data represented by these structures. The second goal of KCODA is to offer better integration of these techniques into the systems of database management by proposing new algorithms allowing to build factorized representations of the data responses to DB requests and by proposing encodings of these representations inside the DB.
ANR Colis — Correctness of Linux Scripts
Participants: Joachim Niehren, Aurélien Lemay, Paul Gallot, Sylvain Salvati.
- Duration: 2015–2021
- Coordinator: R. Treinenm, Université Paris Diderot
- Partner: C. Marché, Tocata project-team, Inria Saclay.
- Objective: This project aims at verifying the correctness of transformations on data trees defined by shell scripts for Linux software installation. The data trees here are the instance of the file system which are changed by installation scripts.
ANR DataCert
Participants: Iovka Boneva, Sophie Tison, Jose Martin Lozano.
- Duration: 2015–2021
- Coordinator: E. Contejean, Université Paris-Sud
- Partners: Université de Lyon
- Objective: The main goals of the DataCert project are to provide deep specification in Coq of algorithms for data integration and exchange and of algorithms for enforcing security policies, as well as to design data integration methods for data models beyond the relational data model.
ANR Headwork
Participants: Joachim Niehren, Momar Ndiouga Sakho, Nicolas Crosetti, Florent Capelli.
- Duration: 2016–2022
- Coordinator: D. Gross-Amblard, Druid Team, Université de Rennes 1
- Scientific partners: Dahu project-team (Inria Saclay) and Sumo project-team (Inria Bretagne)
- Industrial partners: Spipoll and Foulefactory.
- Objective: The main object is to develop data-centric workflows for programming crowd-sourcing systems in flexible declarative manner. The problem of crowd-sourcing systems is to fill a database with knowledge gathered by thousands or more human participants. A particular focus is to be put on the aspects of data uncertainty and for the representation of user expertise.
ANR Delta
Participants: Joachim Niehren, Sylvain Salvati, Aurélien Lemay.
- Duration: 2016–2021
- Partners: LIF (Université Aix-Marseille) and IRIF (Université Paris-Diderot)
- Coordinator: M. Zeitoun, LaBRI (Université de Bordeaux)
- Objective: Delta is focused on the study of logic, transducers and automata. In particular, it aims at extending classical framework to handle input/output, quantities and data.
ANR Bravas
Participants: Sylvain Salvati.
- Duration: 2017–2022
- Coordinator: Jérôme Leroux, LaBRI, Université de Bordeaux
- Scientific Partner: LSV, ENS Cachan
- Objective: The goal of the BraVAS project is to develop a new and powerful approach to decide the reachability problems for Vector Addition Systems (VAS) extensions and to analyze their complexity. The ambition here is to crack with a single hammer (ideals over well-orders) several long-lasting open problems that have all been identified as a barrier in different areas, but that are in fact closely related when seen as reachability.
8.4 Regional initiatives
Dynamic Semantic Crossords, a project of CPER Data
Participants: Joachim Niehren, Chérif Amadou Ba.
- Duration: 2020–2021
- Objective: The objective is to integrate streaming algorithms into the Links demonstrator of dynamic semantic networks.
Knowledge Compilation, a cooperation with Lens, CPER Data
Participants: Florent Capelli.
- Duration: 2020–2021
- F. Capelli cooperates on knowledge compilation with J.-M. Lagniez et P. Marquis. A joined paper got published at AAAI 2021 6. This cooperation is partially funded by the CPER Data.
CPER Cornelia on Artificial Intelligence
Participants: Joachim Niehren.
- Duration: 2021–2025
- The whole Links project is partner of this new CPER project.
PhD project Nicolas Crosetti
Participants: Sophie Tison, Florent Capelli, Joachim Niehren.
- Duration: since 2018
- Cofunded by the Hauts-de-France region. In coopertion with Jan Ramon from Inria Magnet.
9 Dissemination
9.1 Promoting Scientific Activities
9.1.1 Scientific Events: Organisation
- Capelli co-organisation of working group Alga (Automata, Logic, Games & Algebra) of the GDR IM of the CNRS
- Capelli co-organisateur of Working group IMIA (Informatique Mathématique Intelligence Artificielle) of the GDR IM of the CNRS.
Member of the Organizing Committees
-
Capelli
Summer School and Workshop Kocoon on Knowledge Compilation. Organised with Marquis and Mengel from Lens. Cancelled for Corona.
More info at: http://
kocoon. gforge. inria. fr/.
9.1.2 Scientific Events: Selection
Member of the Conference Program Committees
- Capelli Program Committee of the 35th AAAI Conference on Artificial Intelligence (AAAI 2021)
- Capelli Program Committee of the International Joint Conference on Artificial Intelligence (IJCAI 2021).
- Capelli Program Committee of SAT 2020.
- Monet Program Committee of the 35th AAAI Conference on Artificial Intelligence (AAAI 2021)
- Niehren Program Committee of the 8th International Conference on Computational Methods in Systems Biology (CMSB 2020).
- Staworko Program Committee of the 23th International Conference on Extending Database Technology (EDBT 2020)
9.1.3 Journal
Member of the Editorial Boards
- Niehren Editorial Board of Fundamenta Informaticae
- Tison Editorial Board of RAIRO-ITA
- Salvati Managing Editor of the Journal JLLI (Springer)
9.1.4 Scientific Expertise
- Salvati Member of Inria's Evaluation Committee
- Tison Vice-president of the Force Awards association
- Stawarko Academic member of Linked Data Benchmark Council (LDBC)
- Staworko Member of Work Group on Property Graph Schemas (standardisation effort)
9.1.5 Research Administration
- Tison Member of the coordinating team of I-SITE Université de Lille – Nord Europe
- Tison Elected member of the Administrative Council of Université de Lille.
- Tison Elected member of the EATCS council (European Association for Theoretical Computer Science)
- Tison Elected member of CNU 27
- Salvati Member of the recruitement committee of the Computer Science department of Université de Lille and of the CRIStAL laboratory.
9.2 Teaching - Supervision - Juries
9.2.1 Teaching Responibilities
- Salvati co-director of the master “MIAGE FA”
- Salvati director of “Licence Informatique-Mathématique”, Université de Lille
- Salvati co-responsible for the program “Parcours renforcé recherche” of “Licence d’Informatique”, Université de Lille
- Salvati member of the departement council, FIL, Université de Lille
- Paperman responsibility of the program “WebAnalyste” of the master MIASH, Université de Lille
- Tison member of the selection board for “Capes” in computer science
- Staworko Coordinator of International Relationships at the Department of Computer Science, Université de Lille
- Capelli responsibile for the first year of “Licence”, UFR LEA, Université de Lille
- Capelli elected member of the council of the UFR LEA, Université de Lille
- Capelli responsible of the program “ParcourSup” of UFR LEA, Université de Lille
- Capelli tester and corrector of entrance exams for the ENS Computer Science and Mathematics programme, 2020
9.2.2 Teaching Activities
- Boneva teaches computer science at DUT Informatique of Université de Lille
- Capelli teaches computer science at UFR LEA of Université de Lille for around 200h per year (Licence and Master). He is also responsible for remediation of Licence 1 at this UFR.
- Lemay teaches computer science at UFR LEA of Université de Lille for around 200h per year (Licence and Master). He is also responsible for computer science and numeric correspondent for its UFR.
- Niehren gives lessons for the 2nd year students of the Master MOCAD (Université de Lille): on information extraction (21h).
- Paperman teaches computer science for a total of around 200h per year. He gives lessons in UFR MISASH (Université de Lille), in Licence and Master. He also gives a database lesson of 25h in Master MOCAD (Université de Lille).
- Salvati teaches computer science for a total of around 230h per year in computer science departement of Université de Lille. That includes Introduction to Computer Science (L1, 50h), Logic (L3, 50h), Algorithmic and operational research (L3, 36h), Functional Programming (L3, 35h), Research Option (L3, 10h), Semantic Web (M2, 30h), Advanced Databases (M1, 20h).
- Staworko teaches computer science for a total of around 200h in UFR MIME (Université de Lille).
- Tison teaches computer science for a total of around 120h at the Université de Lille. That includes a course on Advanced Algorithms and Complexity (50h, M1), Business Intelligence (36h, M1), Databases (21h L2).
9.2.3 Supervision
- Sakho PhD thesis defended in July. “Certain Query Answering on Hyperstreams” 29. Supervised by Niehren and Boneva.
- Lozano PhD thesis defended in December. “Data Exchange from Relational Databases to RDF with Target Shape Schemas” 28. Supervised by Staworko and Boneva.
- Gallot PhD project in progress since 2017. “On safety of data transformations”. Supervised by Salvati and Lemay.
- Crosetti PhD project in progress since 2018. “Privacy Risks of Aggregates in Data Centric-Workflows”. Supervised by Tison, Capelli, Niehren. With Jan Ramon from Inria Magnet.
- Soyez-Martin PhD project started 2020. “On Streaming with vectors and circuits”. Supervised by Salvati and Paperman.
- Al Serhaly PhD project started 2020. “On hyperstream programming”. Supervised by Niehren.
9.2.4 Juries
PhDs committees
- Tison Membre du jury de thèse de Théo Grente (Caen)
- Tison Membre du jury de thèse de Alexandre Mansard (La Réunion),
- Tison Membre du jury de thèse de Mohammed Houssem Eddine Hachmaoui (Saclay, présidente du jury)
HDR committees
- Niehren Reviewer of the HDR of Loïc Paulevé, Université Paris-Saclay.
- Salvati Member of the jury for the HDR of Olivier Gauwin, Université de Bordeaux.
10 Scientific production
10.1 Major publications
- 1 inproceedings Topological Sorting with Regular Constraints 45th International Colloquium on Automata, Languages, and Programming (ICALP 2018) Prague, Czech Republic July 2018
- 2 inproceedings The Tractability of SHAP-Score-Based Explanations over Deterministic and Decomposable Boolean Circuits Thirty-Fifth AAAI Conference on Artificial Intelligence Held online, France February 2021
- 3 inproceedings Stackless Processing of Streamed Trees 2021 PODS Xi'an, Shaanx, China June 2021
- 4 inproceedings Semantics and Validation of Shapes Schemas for RDF ISWC2017 - 16th International semantic web conference Vienna, Austria October 2017
- 5 inproceedings Oblivious and Semi-Oblivious Boundedness for Existential Rules IJCAI 2019 - International Joint Conference on Artificial Intelligence Macao, China August 2019
- 6 inproceedings Certifying Top-Down Decision-DNNF Compilers Thirty-Fifth AAAI Conference on Artificial Intelligence Online, France February 2021
- 7 inproceedings Tractable QBF by Knowledge Compilation 36th International Symposium on Theoretical Aspects of Computer Science (STACS 2019) https://arxiv.org/abs/1807.04263 Berlin, Germany March 2019
- 8 inproceedings Linear high-order deterministic tree transducers with regular look-ahead MFCS 2020 : The 45th International Symposium on Mathematical Foundations of Computer Science Andreas Feldmann and Michal Koucky and Anna Kotesovcova Prague, Czech Republic INRIA August 2020
- 9 article Determinization and Minimization of Automata for Nested Words Revisited Algorithms February 2021
- 10 inproceedings Containment of Shape Expression Schemas for RDF SIGMOD-SIGACT-SIGAI Symposium on Principles of Database Systems (PODS) Amsterdam, Netherlands June 2019
10.2 Publications of the year
International journals
International peer-reviewed conferences
National peer-reviewed Conferences
Scientific book chapters
Doctoral dissertations and habilitation theses
Reports & preprints