Magnet is a new research group that aims to design new machine learning based methods geared towards mining information networks. Information networks are large collections of interconnected data and documents like citation networks and blog networks among others. We will then define new prediction methods for texts and networks of texts based on machine learning algorithms in graphs. Such algorithms include node and link classification, link prediction, clustering and probabilistic modeling of graphs. Envisioned applications include browsing, monitoring and recommender systems, and more broadly information extraction in information networks. Application domains cover social networks for cultural data and e-commerce, and biomedical informatics.

The main objective of Magnet is to develop original machine learning methods for networked data. We consider information networks in which the data are vectorial data and texts. We model such information networks as (multiple) (hyper)graphs wherein nodes correspond to entities (documents, spans of text, users, ...) and edges correspond to relations between entities (similarity, answer, co-authoring, friendship, ...). Our main research goal is to propose new learning algorithms to build applications like browsing, monitoring and recommender systems, and more broadly information extraction in information networks. Hence, we will investigate new learning algorithms for node clustering and node classification, link classification and link prediction. Also, we will search for the best hidden graph structure to be generated for solving a given learning task. We will base our research on generative models for graphs, on machine learning for graphs and on machine learning for texts. The challenges are the dimensionality of the input space, possibly the dimensionality of the output space, the high level of dependencies between the data, the inherent ambiguity of textual data and the limited amount of human labeling. An additional challenge will be to design scalable methods for large information networks. Hence, we will explore how sampling and randomization can be used in new machine learning algorithms. Also, active machine learning algorithms for graphs will be investigated.

On the first hand we want to design machine learning algorithms on graphs to solve problems in networks of texts and documents in natural language. The main originality of this research is to consider and take advantage of the setting of networked data exploiting the relationships between different data entities and, overall, the graph topology. On the second hand, in a concomitant way, we want to develop prediction models for graph-like data. This includes prediction, ranking and classification of links and nodes in an on-line or batch setting. The two objectives are intertwined, enrich each other and raise important scientific questions we want to focus on. Our research proposal is organized according to the following questions:

How to go beyond vectorial classification models in natural language oriented tasks?

How to adaptively build graphs with respect to the given tasks? How to create network from observations of information diffusion processes?

How to design methods able to achieve very good predictive accuracy without giving up on scalability?

How to go beyond strict node homophilic/similarity assumptions in graph-based learning methods?

One of our overall research objectives is to derive graph-based machine learning algorithms for natural language and text information extraction tasks. This section discusses the motivations behind the use of graph-based ML approaches for these tasks, the main challenges associated with it, as well as some concrete projects. Some of the challenges go beyond NLP problems and will be further developed in the next sections. An interesting aspect of the project is that we anticipate some important cross-fertilizations between NLP and ML graph-based techniques, with NLP not only benefiting from but also pushing ML graph-based approaches into new directions.

Motivations for resorting to graph-based algorithms for texts are at least threefold. First, online texts are organized in networks. With the advent of the web, and the development of forums, blogs, and micro-blogging, and other forms of social media, text productions have become strongly connected. Thus, documents on the web are linked through hyperlinks, forum posts and emails are organized in threads, tweets can be retweeted, etc. Additional connections can be made through users connections (co-authorship, friendship, follower, etc.). Interestingly, NLP research has been rather slow in coming to terms with this situation, and most work still focus on document-based or sentence-based predictions (wherein inter-document or inter-sentence structure is not exploited). Furthermore, several multi-document tasks exist in NLP (such as multi-document summarization and cross-document coreference resolution), but most existing work typically ignore document boundaries and simply apply a document-based approach, therefore failing to take advantage of the multi-document dimension , .

A second motivation comes from the fact that most (if not all) NLP
problems can be naturally conceived as graph problems. Thus, NL tasks
often involve discovering a relational structure over a set of text
spans (words, phrases, clauses, sentences, etc.). Furthermore, the
*input* of numerous NLP tasks is also a graph; indeed, most
end-to-end NLP systems are conceived as pipelines wherein the output
of one processor is in the input of the next. For instance, several
tasks take POS tagged sequences or dependency trees as input. But this
structured input is often converted to a vectorial form, which
inevitably involves a loss of information.

Finally, graph-based representations and learning methods in principle
appear to address some core problems faced by NLP, such as the fact
that textual data are typically not independent and identically
distributed, they often live on a manifold, they involve very high
dimensionality, and their annotations is costly and scarce. As such,
graph-based methods represent an interesting alternative, or at least
complement, to structured prediction methods (such as CRFs or
structured SVMs) commonly used within NLP. While structured output
approaches are able to model local dependencies (e.g., between
neighboring words or sentences), they cannot efficiently capture long
distance dependencies, like forcing a particular

Given the above motivations, our first line of research will be to investigate how one can leverage an underlying network structure (e.g., hyperlinks, user links) between documents, or text spans in general, to enhance prediction performances for several NL tasks. We think that a “network effect”, similar to the one that took place in Information Retrieval (with the Page Rank algorithm), could also positively impact NLP research. A few recent papers have already opened the way, for instance in attempting to exploit Twitter follower graph to improve sentiment classification .

Part of the challenge in this work will be to investigate how adequately and efficiently one can model these problems as instances of more general graph-based problems, such as node clustering/classification or link prediction discussed in the next sections. In a few cases, like text classification or sentiment analysis, graph modeling appears to be straightforward: nodes correspond to texts (and potentially users), and edges are given by relationships like hyperlinks, co-authorship, frienship, or thread membership. Unfortunately, modeling NL problems as networks is not always that obvious. From the one hand, the right level of representation will probably vary depending on the task at hand: the nodes will be sentences, phrases, words, etc. From the other hand, the underlying graph will typically not be given a priori, which in turn raises the question of how we construct it. Of course, there are various well-known ways to obtain similarity measures between text contents (and its associated vectorial data), and graphs can be easily constructed from those combined with some sparsification method. But we would like our similarity to be tailored to the task objective. An additional problem with many NLP problems is that features typically live in different types of spaces (e.g., binary, discrete, continuous). A preliminary discussion of the issue of optimal graph construction for semi-supervised learning in NLP is given in , . We identify the issue of adaptative graph construction as an important scientific challenge for machine learning on graphs in general, and we will discuss it further in Section .

As noted above, many NLP tasks have been recast as structure
prediction problems, allowing to capture (some of the) output
dependencies. Structure prediction can be viewed as (set of) link
prediction with global loss or dependencies, which means that
graph-based learning methods can handle (at least, approximately)
output prediction dependencies, and they can in principle capture
additional more global dependencies given the right graph
structure. How to best combine structured output and graph-based ML
approaches is another challenge that we intend to address. We will
initially investigate this question within a semi-supervised context,
concentrating on graph based regularization and graph propagation
methods. Within such approaches, labels are typically binary or they
correspond to small finite set. Our objective is to explore how one
propagates an exponential number of *structured labels* (like a
sequence of tags or a dependency tree) through graphs. Recent attempts
at blending structured output models with graph-based models are
investigated in , . Another
related question that we will address in this context is how does one
learn with *partial labels* (like partially specificied tag
sequence or tree) and use the graph structure to complete the output
structure. This last question is very relevant to NL problems where
human annotations are costly; being able to learn from partial
annotations could therefore allow for more targeted annotations and in
turn reduced costs .

The NL tasks we will mostly focus on are coreference resolution and entity linking, temporal structure prediction, and discourse parsing. These tasks will be envisonned in both document and cross-document settings, although we expect to exploit inter-document links either way. Choices for these particular tasks is guided by the fact that are still open problems for the NLP community, they potentially have a high impact for industrial applications (like information retrieval, question answering, etc.), and we already have some expertise on these tasks in the team. As a midterm goal, we also plan to work on tasks more directly relating to micro-blogging, such sentiment analysis and the automatic thread structuring of technical forums; the latter task is in fact an instance of rhetorical structure prediction .

We have already initiated some work on the coreference resolution
problem in the context of ML graph-based approaches. We cast this
problem as a spectral clustering problem. Given than features can be
numerical or nominal, the definition of a good similarity measure
between entities is not straightforward. As a first solution, we
consider only numerical attributes to build a

In most applications, edge weights are computed through a complex data-modeling process and convey crucially important information for classifying nodes, which makes it possible to infer information related to each data sample even exploiting the graph topology solely. In fact, a widespread approach to the solution of several classification problems is representing the data through an undirected weighted graph in which edge weights quantify the similarity between data points. This technique for coding input data has been applied to several domains, including classification of genomic data ( ), face recognition ( ), and text categorization ( ).

In some cases, the full adjacency matrix is generated by employing
suitable similarity functions chosen through a deep understanding of
the problem structure. For example TF-IDF representation of documents,
the affinity between pairs of samples is often estimated through the
cosine measure or the

In this project we will address the problem of adaptive graph
construction towards several directions. One is the question of
choosing the best similarity measure given the objective learning
task. This question is related to the question of similarity learning
( ) which has not been considered in the
context of graph based learning. In the context of structured
prediction, we will develop approaches where output structures are
organized in graphs whose similarity is given by top-

A different way we envision adaptative graph construction is in the context of semi-supervised learning. Partial supervision can take various forms and an interesting and original setting is governed by two currently studied applications: detection of brain anomaly from connectome data and polls recommendation in marketing. Indeed, for these two applications, a partial knowledge of the information diffusion process can be observed while the network is unknown or only partially known. An objective is to construct (or complete) the network structure from some local diffusion information. The problem can be formalized as a graph construction problem from partially observed diffusion processes. It has been studied very recently in . In our case, the originality comes either from the existence of different sources of observations or from the large impact of node contents in the network.

We will study how to combine graphs defined by networked data and graphs built from flat data to solve a given task. This is of major importance for information networks because, as said above, we will have to deal with multiple relations between entities (texts, spans of texts, ...) and also use textual data and vectorial data. We have started to work on combining graphs in a semi supervised setting for node classification problems along the PhD thesis of T. Ricatte. Future work include combination geared by semi-supervision on link prediction tasks. This can be studied in an active learning setting. But one important issue is to design scalable approaches, thus to exploit locality given by the network. Doing this we address another objective to build non uniformly parameterized combinations.

As stated in the previous sections, graphs as complex objects provides a rich representation of data. Often enough the data is only partially available and the graph representation is very helpful in predicting the unobserved elements. We are interested in problems where the complete structure of the graph needs to be recover and only a fraction of the links is observed. The link prediction problem falls into this category. We are also interested in the recommendation and link classification problems which can be seen as graphs where the structure is complete but some labels on the links (weights or signs) are missing. Finally we are also interested in labelling the nodes of the graph, with class or cluster memberships or with a real value, provided that we have (some information about) the labels for some of the nodes.

The semi-supervised framework will be also considered. A midterm research plan is to study how graph-based regularization models help for structured prediction problems. This question will be studied in the context of NLP tasks, as noted in Section , but we also plan to develop original machine learning algorithms that have a more general applicability. Inputs are networks whose nodes (texts) have to be labeled by structures. We assume that structures lie in some manifold and we want to study how labels can propagate in the network. One approach is to find smooth labeling function corresponding to an harmonic function on both manifolds in input and output. We also plan to extend our results on spectral clustering with must-link and cannot-link constraints in two directions. We have proposed a batch method with an optimization problem based on an adaptive spectral embedding with respects to constraints. We want to extend this approach to an on-line and active setting where a flow of graphs (each one is a document) is given as input. In the case of large graphs, we also consider the case where partial supervision consists in the knowledge of few clusters.

Scalability is one of the main issue in the design of new prediction algorithms working on networked data. It has gained more and more importance in recent years, because of the growing size of the most popular networked data that are now used by millions of people. In such contexts, learning algorithms whose computation time scales quadratically, or slower, in the number of considered data objects (usually nodes or vertices, depending on the given task) should be considered impractical.

These observations lead to the idea of using graph sparsification techniques in order to work on a part of the original network for getting results that can be easily extended and used for the whole original input. A sparsified version of the original graph can often be seen as a subset of the initial input, i.e. a suitably selected input subgraph which forms the training set (or, more in general, it is included in the training set). This holds even for the active setting.

A simple example could be to find a spanning tree of the input graph, possibly using randomization techniques, with properties such that we are allowed to obtain interesting results for the initial graph dataset. We have started to explore this research direction for instance in . This approach leaves us with the problem of choosing a good spanning tree, taking into account that the setting could be adversarial (e.g, in the online case the presentation and the assignment of the labels are both arbitrary). A suitable use of the randomization power becomes therefore remarkably significant. Moreover, it is interesting to observe that running a prediction algorithm on a sparsified version of the input dataset allows the parallelization of prediction tasks. In fact, given a prediction task for a networked dataset, in a preliminary phase one could run a randomized graph sparsification method in parallel on different machines. For example, in the case of the spanning tree use, one could then draw several spanning trees at the same time, each on a different computer. This way it is possible to simultaneously run different prediction experiments on the same task and aggregating the obtained results at the end, with several methods (e.g. simply by majority vote) in order to increase the robustness and accuracy predictions.

At the level of the mathematical foundations, the key issue to be addressed in the study of (large-scale) random networks also concerns the segmentation of network data into sets of independent and identically distributed observations. If we identify the data sample with the whole network, as it has been done in previous approaches , we typically end up with a set of observations (such as nodes or edges) which are highly interdependent and hence overly violate the classic i.i.d. assumption. In this case, the data scale can be so large and the range of correlations can be so wide, that the cost of taking into account the whole data and their dependencies is typically prohibitive. On the contrary, if we focus instead on a set of subgraphs independently drawn from a (virtually infinite) target network, we come up with a set of independent and identically distributed observations—namely the subgraphs themselves, where subgraph sampling is the underlying ergodic process . Such an approach is one principled direction for giving novel statistical foundations to random network modeling. At the same time, because one shifts the focus from the whole network to a set of subgraphs, complexity issues can be restricted to the number of subgraphs and their size. The latter quantities can be controlled much more easily than the overall network size and dependence relationships, thus allowing to tackle scalability challenges through a radically redesigned approach.

We intend to develop new learning models for link prediction problems. We have already proposed a conditional model in with statistics based on Fiedler values computed on small subgraphs. We will investigate the use of such a conditional model for link prediction. We will also extend the conditional probabilistic models to the case of graphs with textual and vectorial data by defining joint conditional models. Indeed, an important challenge for information networks is to introduce node contents in link ranking and link prediction methods that usually rely solely on the graph structure. A first step in this direction was already proposed in where we learn a mapping of node content to a new representation constrained by the existing link structure and applied it for link recommendation. This approach opens a different view on recommendation by means of link ranking problems for which we think that non parametric approaches should be fruitful.

Regarding link classification problems, we plan to devise a whole family of active learning strategies, which could be based on spanning trees or sparse input subgraphs, that exploit randomization and the structure of the graph in order to offset the adversarial label assignment. We expect these active strategies to exhibit good accuracies with a remarkably small number of queried edges, where passive learning methods typically break down. The theoretical findings can be supported by experiments run on both synthetic and real-world (Slashdot, Epinions, Wikipedia, and others) datasets.

We are interested in studying generative models for graph labeling, exploiting the results obtained in p-stochastic model for link classification (investigated in ) and statistical model for node label assignment which can be related to tree-structured Markov random fields .

In developing our algorithms, we focus on providing theoretical guarantees on prediction accuracy and, at the same time, on computational efficiency. The development of methods that simultaneously guarantee optimal accuracy and computational efficiency is a very challenging goal. In fact, the accuracy of most methods in the literature is not rigorously analyzed from a theoretical point of view. Likewise, tight time and space complexity bounds are not generally provided. This contrasts with the need to manage extremely large relational datasets like, e.g., snapshots of the World Wide Web.

In many cases, the algorithms devised for solving node classification problems are driven by the following assumption: linked entities tend to be assigned to the same class. This assumption, in the context of social networks, is known as homophily ( , ) and involves ties of every type, including friendship, work, marriage, age, gender, and so on. In social networks, homophily naturally implies that a set of individuals can be parted into subpopulations that are more cohesive. In fact, the presence of homogeneous groups sharing interests is one of the most significant reasons for affinity among interconnected individuals, which suggests that, in spite of its simplicity, this principle turns out to be very powerful for node classification problems in general networks.

Recently, however, researchers have started to consider networked data where connections may also carry a negative meaning. For instance, disapproval or distrust in social networks, negative endorsements on the Web. Concrete examples are provided by certain types of online social networks. Users of Slashdot can tag other users as friends or foes. Similarly, users of Epinions can give positive or negative ratings not only to products but also to other users. Even in the social network of Wikipedia administrators, votes cast by an admin in favor or against the promotion of another admin can be viewed as positive or negative links. More examples of signed links are found in other domains, such as the excitatory or inhibitory interactions between genes or gene products in biological networks.

Although the introduction of signs on graph edges appears like a small change from standard weighted graphs, the resulting mathematical object, called signed graph, has an unexpectedly rich additional complexity. For example, the spectral properties of signed graphs, which essentially all sophisticated node classification algorithms rely on, are different and less known than those of their unsigned counterparts. Signed graphs naturally lead to a specific inference problem that we have discussed in previous sections: link classification. This is the problem of predicting the sign of links in a given graph. In online social networks, this may be viewed as a form of sentiment analysis, since we would like to semantically categorize the relationship between individuals.

To conclude, we plan to go beyond the homophilic bias from the algorithmic as well as from the modeling point of view. We will consider new kind of modeling and learning biases provided by graphs with negative weights (signed graphs) and hypergraphs. We will study their spectral properties, smoothness measures of (node or edge) labeling. Sampling and walking also need to be reconsidered. From the machine learning perspective, we will study edge and node labeling in batch and online settings. In connection with our main targeted applications, we will mainly consider unsupervised and semi-supervised situations. We think that allowing negative weights and advanced relationships on nodes will also lead to space efficient representations of graphs.

Our main targeted applications are browsing, monitoring and mining in information networks. Such discovered structures would also be beneficial to predicting links between users and texts which is at the core of recommender systems. All the learning tasks considered in the project such as node clustering, node and link classification and link prediction are likely to yield important improvements in these applications. Application domains cover social networks for cultural data and e-commerce, and biomedical informatics.

CoRTex is a LGPL-licensed Python library for Noun Phrase coreference resolution in
natural language texts. This library contains implementations of
various state-of-the-art coreference resolution algorithms, including
those developed in our research.
In addition, it provides a set of
APIs and utilities for text pre-processing, reading the main
annotation formats (ACE, CoNLL and MUC), and performing evaluation
based on the main evaluation metrics (MUC, B-CUBED, and CEAF). As
such, CoRTex provides benchmarks for researchers working on
coreference resolution, but it is also of interest for
developers who want to integrate a coreference resolution within a
larger platform. This project is hosted on Inria
gforge: https://

We developed a new framework for high order learning .

We have illustrated the usefulness of automatically annotated examples in complex learning supervised by few training examples , .

We propose a new algorithm for semi-supervised spectral clustering and apply it to the NLP task of noun phrase coreference resolution .

Along the thesis of Thomas Ricatte, in
and , we propose methods for learning from
interactions between groups in networks. We propose a proper extension
of graphs, called hypernode graphs as a formal tool able to model
group interactions. A hypernode graph is a collection of weighted
relations between two groups of distinct nodes. Weights quantify the
individual participation of nodes to a given relation. We define
Laplacians and kernels for hypernode graphs and and prove that they
strictly generalize over graph kernels and hypergraph kernels. We
prove that hypernode graphs correspond to signed graphs such that the
matrix

In and , we develop a new approach for the automatic identification of so-called implicit discourse relations. Specifically, our system combines hand labeled examples and automatically annotated examples based on explicit relations using several simple methods inspired by work in domain adaptation. Our system is evaluated empirically on the Annodis corpus, a French corpus annotated with discourse structures. Our system yields significant performance gains compared to only using hand-labeled data or using only automatically annotated data.

We worked on developing a new algorithm in order to construct a graph in a adaptive way for a specific task. More precisely, we looked for a metric learning algorithm that could depend on the target task. Previous works on metric learning ( ) aim at learning a relevant metric using a linear approach, which cannot capture the non-linearity of the data. Our approach, instead, aims at learning the most appropriate non-linear data projection for the target task. For this purpose, we train a neural network with relative constraints depending on the target task and a target classic metric (e.g. euclidean distance, cosine similarity, ...), in order to make the metric meaningful for the new data representation and our target task.

From a mathematical point of view, signed networks are graphs whose edges carry a sign representing the positive or negative nature of the relationship between the incident nodes. These structures are extremely useful for modeling, at the same time, similarity and dissimilarity object relationships. Given an undirected signed graphs, in the Correlation Clustering problem the goal is to find a node partition into clusters minimizing the number of negative (dissimilarity) edges linking two nodes within the same cluster and the number of positive (similarity) edges between different clusters.

We focused on devising an algorithm able to solve the Correlation Clustering problem for general input signed graphs (if the input is a complete signed graph the problem is proven to be much easier). One of the main objective of this work is the use of the proposed algorithm for creating a learner able to predict the unknown edges signs of a given signed graph. This prediction task is known as Link Classification in signed graphs. In fact, given an undirected signed graph whose edge set is split into training and test set, we could use the Correlation Clustering solution working for general input graphs for partitioning the training set and using the node partition generated for predicting the test edge signs. Moreover, one could exploit such an algorithm for developing new strategies for the Link Classification problem operating within the online and active Machine Learning setting.

Since the node set partitioning turns out to be strictly related to the Link Classification problem, we also focused on the very challenging goal of obtaining a deep understanding of the complex interplay between Link and Node Classification. More precisely, we investigated the relationships between the Vapnik Chervonenkis dimension of any given set of hypothesis space of node and edge similarity functions operating within this framework.

Given a set of objects (vertices of a graph) and a set of pairwise preference labels between objects (directed edges connecting vertices) which may be non-transitive due to irrationality or arbitrary noise, what is a correct way to sample preference labels for ordering the set of objects? This long standing open problem is, as far as we know, unsolved when each pairwise preference labels refers to two (disjoint) sets of objects (vertices). This framework can be easily motivated considering that quite often, in many real world contexts, users express their preferences between sets of items rather than single items, and turns out to be strictly connected with our recent model of hypergraphs with bipartite hyperedges . We are working on devising a new algorithm able to rank a given set of items (graph node set) when only comparisons between sets containing at least 2 items are allowed. This challenging and interested problem is, as far as we know, quite novel and can be studied within different Machine Learning setting (online, batch, active, ...). The preliminaries results we are obtaining, when setting the cardinality of the item sets equal to 2, are encouraging and indicate that it could be possible to extend our strategies in order to deal with larger item sets.

In this section we provide the results we obtained that are not related with our main research directions.

In business intelligence information systems, one of the first tasks is to acquire and clean internal data and then enrich them with additional sources of informations. This preprocessing step is well planned and specialized for fixed analysis and fixed dashboards. The subject of our collaboration with SAP was included in general objective that (i) specializes this preprocessing task in order to deal with external data coming from networked data like social networks and open relational data, and (ii) simplifies the adaptation of the processing step evolving data analysis tasks. We have focused on the task of merging information acquired from many input data sources represented as graphs, with the final objective of providing a unique graph representation of all data or data models. This research has lead to new graph combination algorithms, but has also raised the need for representing and managing high order relations using graph-like techniques.

Rémi Gilleron supervises the PhD thesis (Cifre) of Thomas Ricatte together with Yannick Cras from SAP.

We start a one to one cooperation with the Clic and Walk company that makes marketing surveys by consumers (called clicwalkers). The goal of the company is to understand the community of clicwalkers (40 thousands in one year) and its evolution with two objectives: the first one is to optimize the attribution of surveys to clicwalkers, and the second is to expand company's market to foreign countries. Social data can be obtained from social networks (G+, Facebook, ...) but there is no explicit network to describe the clicwalkers community. But users activity in answering surveys as well as server logs can provide traces of information diffusion, geolocalisation data, temporal data, sponsorship, ...We will study the problem of adaptive graph construction from the clicwalkers network. Node (users) classification and clustering algorithms will be applied. For the problem of survey recommendations, the problem of teams constitution in a bipartite graphs of users and surveys will be studied. Random graph modeling and generative models of random graphs will be one step towards the prediction of the evolution of clicwalkers community.

Mikaela Keller and Marc Tommasi supervise the PhD thesis (Cifre) of Pauline Wauquier on graph-based recommendation together with Guillaume André from Clic and Walk.

Marc Tommasi and Pascal Denis supervise the PhD thesis of David Chatel on semi-supervised spectral clustering. The PhD is funded by Inria and the "Région Nord – Pas de Calais".

Marc Tommasi belongs to the scientific committee involved in the process of building the IDEX proposal for Lille.

The Lampada project on “Learning Algorithms, Models and sPArse
representations for structured DAta” is coordinated by Tommasi from
Mostrare. Our partners are the Sequel project of Inria Lille
Nord Europe, the Lif (Marseille), the Hubert Curien
laboratory (Saint-Etienne), and lip6 (Paris). More information
on the project can be found on
http://

We are part of FUI Hermes (2012-2015), a joint project in collaboration with many companies (Auchan, KeyneSoft, Cylande, ...). The main objective is to develop a platform for contextual customer relation management. The project started in November 2012.

Pascal Denis is an associate member of the Laboratoire d'Excellence *Empirical Foundations of Linguistics* (EFL), http://

Fabien Torre is elected for "CNU section 27 (informatique)" since Oct. 2011

Program: ERC Advanced Grant

Project acronym: STAC

Project title: Strategic conversation

Duration: Sept. 2011 - Aug. 2016

Coordinator: Nicholas Asher, CNRS, Université Paul Sabatier, IRIT (France)

Other partners: School of Informatics, Edinburgh University; Heriot Watt University, Edinburgh

Abstract: STAC is a five year interdisciplinary project that aims to develop a new, formal and robust model of conversation, drawing from ideas in linguistics, philosophy, computer science and economics. The project brings a state of the art, linguistic theory of discourse interpretation together with a sophisticated view of agent interaction and strategic decision making, taking advantage of work on game theory.

We invited Prof. Claudio Gentile (University of Insubria, Italy) in January (he gave a talk on "Online Clustering of Bandits in a Social Network") and in June.

Prof. Nicolò Cesa-Bianchi (University of Milan, Italy) visited us in July (he gave a talk on "The Online Approach to Machine Learning").

Finally, we invited Prof. Mark Herbster (University College London) in July (he gave a talk on "Online Approximate Prediction at the Limit of Zero Temperature in an Ising Model") and November.

In May Fabio Vitale visited the Department of Computer Science of the University of Milan, collaborating with Prof. Nicolò Cesa-Bianchi.

Pascal Denis served as member of the program committee of ACL 2014, COLING 2014, EACL 2014, and EMNLP 2014.

Rémi Gilleron served as member of the program committee of CAp'2014.

Marc Tommasi served as member of the program committee of CAp'2014.

Fabien Torre served as member of the program committee of ICPRAM 2015, EGC 2015, CluCo 2015, ECAI 2014, ICPRAM 2014.

Mikaela Keller served as member of the program committee of NIPS 2014 and CAp'2014, of the scientific committee reviewing Digiteo post-doctoral grant proposal.

Mikaela Keller was also reviewer for EGC 2015.

Fabio Vitale was reviewer for NIPS 2014.

Master MOCAD: Pascal Denis, Extraction d'information, 18h, M2, Université Lille 1, France

Master MOCAD: David Chatel, Extraction d'information, 3h30, M2, Université Lille 1, France

Master MIASHS: Rémi Gilleron, Bases de données avancées, 24h, M1, Université de Lille, France

Master MIASHS: Rémi Gilleron, Programmation R (1), 24h, M1, Université de Lille, France

Master MIASHS: Rémi Gilleron, Programmation R (2), 24h, M1, Université de Lille, France

Master MIASHS: Rémi Gilleron, Algorithmes de fouille de données, 36h, M1, Université de Lille, France

Licence MIASHS: Rémi Gilleron, Programmation Python, 36h, L1, Université de Lille, France

Licence : Mikaela Keller, Introduction to computer science and programming, 46h, L1 and L2, Université Lille 3, France

Licence : Mikaela Keller, Introduction to Statistical Learning, 54h, L3, Université Lille 3, France

Licence : Mikaela Keller, C2i, 55h, L2-L3, Université Lille 3, France

Licence MIASHS: Marc Tommasi, Réseaux, 64h, L1 and L2, Université de Lille, France

Licence MIASHS: Marc Tommasi, Programmation client, 24h, L2, Université de Lille, France

Licence : Marc Tommasi, C2i, 25h, L2, Université de Lille, France

Licence : Marc Tommasi, Culture numérique, 30h, L2, Université de Lille, France

Master : Fabien Torre, langages statiques du web, 37h, M1, Université Lille 3, France

Master : Fabien Torre, algorithmique et programmation php pour le web, 75h, M1, Université Lille 3, France

Master : Fabien Torre, algorithmique et programmation pour l'extraction d'information, 55h, M2, Université Lille 3, France

Master : Fabien Torre, javascript langage dynamique du web, 38h, M2, Université Lille 3, France

Licence : Fabio Vitale, C2i, 25h, L1, Université Lille 3, France

Licence MIASHS (SHS) : Fabio Vitale, Intelligence Artificielle, 31h30, L2, Université Lille 3, France

Master MIASHS (SHS) : Fabio Vitale, Classification non supervisée, 30h, M1, Université Lille 3, France

Licence SHDS (SHS) : Fabio Vitale, TICE, 18h, L1, Université Lille 3, France

Licence SQ (SHS) : Fabio Vitale, Algorithmique de graphes, 19h, L3, Université Lille 3, France

Fabien Torre teaches a course at EPAT 2014 (École de Printemps sur l'Apprentissage arTificiel) on Machine Learning Collaborative Methodes.

Rémi Gilleron is in charge of Master Mathématiques et Informatique Appliquées aux Sciences Humaines et Sociales (MIASHS) at Université de Lille.

Marc Tommasi is a member of the "conseil de l'UFR MIME".

Fabien Torre is "secrétaire du CHSCT (Comité hygiène, sécurité et conditions de travail) de Lille 3" since Dec. 2012.

PhD: Thomas Ricatte, Hypernode graphs : a new data structure for higher order learning, Université Lille 3, since May 2011, defense scheduled for Jan. 2015, Rémi Gilleron

PhD in progress: David Chatel, Supervised Spectral Clustering and Information Diffusion in Graphs of Texts, Université Lille 1, since Sept. 2012, Pascal Denis and Marc Tommasi

PhD in progress: Pauline Wauquier, Recommendation in Information Networks, Université Lille 1, since Dec. 2013 supervised by Marc Tommasi and Mikaela Keller

PhD: Jean Decoster, Programmation logique inductive pour la classification et la transformation de documents semi-structurés, Université Lille 1, since Sept. 2009, defense on July 2014, Fabien Torre, Rémi Gilleron and Slawomir Staworko

PhD in progress: Emmanuel Lassalle, Improved Coreference Resolution with Feature Space Learning, Université Paris-Diderot, since Sept. 2010, Pascal Denis and Laurence Danlos (Université Paris-Diderot)

PhD in progress: Chloé Braud, Discourse Relation Identification from Labeled and Unlabeled Data, Université Paris-Diderot, since Sept. 2011, Pascal Denis and Laurence Danlos (Université Paris-Diderot)

Master in progress: Mathieu Dehouck, Propagating structured labels for Dependency Parsing, Université Lille 1, supervised by Pascal Denis

PhD in progress: Géraud Le Falher, Machine Learning in Signed Graphs, Inria Lille – Nord Europe, since Oct. 2014, Marc Tommasi, Fabio Vitale and Claudio Gentile (University of Insubria, Italy).

Marc Tommasi wrote a report for the Phd thesis of Daniel Bernades ("LIP6", France).

Mikaela Keller was a member of the following selection/hiring committees: PhD committee of J.P. Peyrache, Université de St Etienne (France), MdC hiring committees at Université Paris-Sud, and MdC hiring committees at Université Paris 6 (France).

Marc Tommasi was member of the following selection/hiring committees: UPMC (Paris 6): Professor, St Etienne (France): Assistant professor and Lille (France): Inria young researchers (CR2).

Marc Tommasi was also reviewer for Crédit Impot Recherche.

Fabien Torre participated in the selection committees "COS" for the 27 MCF 0133 and 27 MCF 0470 position (spring 2014, Lens, France).

Marc Tommasi was invited to give seminars at (i) Inria Rennes and (ii) LiX Colloquium on data science.

Mikaela Keller was involved in "Fête de la science" as a "chercheur itinérant".