Search

Soumission à Episciences, Mobile ad-hoc social networks (MASNs) have been the subject of several research studies over the past two decades. They allow stations located in a small geographical area to be connected without the need for a network infrastructure and offer them the possibility to communicate any time anywhere. To communicate, stations regularly broadcast their interests in the form of keywords. Stations with a high degree of similarity among their keywords can communicate with each other. However, the coverage of MASNs is limited to a small geographical area, due to the limited communication range of mobile ad-hoc networks (MANET) stations. In this paper, we present an architecture and implementation of hybrid mobile ad-hoc social networks (MASNs coupled to infrastructure networks) of Android mobile devices for information dissemination. Stations can use the infrastructure network to communicate and rely on the mobile ad-hoc network when the infrastructure is not available.Rather than communicating synchronously as this is the case in the similar works found in the literature, in our approach, the stations communicate using a publish/subscribe communication protocol, which is perfectly suited to this type of network thanks to the decoupling in time and space it provides., Les réseaux sociaux mobiles ad hoc (MASN) ont fait l'objet de plusieurs recherches au cours des deux dernières décennies. Ils permettent à des stations situées dans une petite zone géographique d'être connectées sans avoir besoin d'une infrastructure de réseau et leur offrent la possibilité de communiquer à tout moment et en tout lieu. Pour communiquer, les stations diffusent régulièrement leurs centres d'intérêt sous forme de mots-clés. Les stations dont les mots-clés présentent un degré élevé de similitude peuvent communiquer entre elles. Cependant, la couverture des réseaux MASN est limitée à une petite zone géographique, en raison de la portée de communication limitée des stations des réseaux mobiles ad hoc (MANET). Dans cet article, nous présentons une architecture et une mise en œuvre de réseaux sociaux mobiles ad hoc hybrides (MASN couplés à des réseaux d'infrastructure) d'appareils mobiles Android pour la diffusion d'informations. Les stations peuvent utiliser le réseau d'infrastructure pour communiquer et s'appuyer sur le réseau mobile ad hoc lorsque l'infrastructure n'est pas disponible.Plutôt que de communiquer de manière synchrone comme c'est le cas dans les travaux similaires trouvés dans la littérature, dans notre approche, les stations communiquent en utilisant un protocole de communication de type publication/abonnement, qui est parfaitement adapté à ce type de réseau grâce au découplage dans le temps et l'espace qu'il fournit.

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

While the ever-increasing emergence of online services has led to a growing interest in the development of recommender systems, the algorithms underpinning such systems have begun to be criticized for their role in limiting the variety of content exposed to users. In this context, the notion of diversity has been proposed as a way of mitigating the side effects resulting from the specialization of recommender systems. In this paper, using a well-known recommender system that makes use of collaborative filtering in the context of musical content, we analyze the diversity of recommendations generated through the lens of the recently proposed information network diversity measure. The results of our study offer significant insights into the effect of algorithmic recommendations. On the one hand, we show that the musical selections of a large proportion of users are diversified as a result of the recommendations. On the other hand, however, such improvements do not benefit all users. They are in fact mainly restricted to users with a low level of activity or whose past musical listening selections are very narrow. Through more in-depth investigations, we also discovered that while recommendations generally increase the variety of the songs recommended to users, they nonetheless fail to provide a balanced exposure to the different related categories.

International audience; Online social platforms have become central in the political debate. In this context, the existence of echo chambers is a problem of primary relevance. These clusters of like-minded individuals tend to reinforce prior beliefs, elicit animosity towards others and aggravate the spread of misinformation. We study this phenomenon on a Twitter dataset related to the 2017 French presidential elections and propose a method to tackle it with content recommendations. We use a quadratic program to find optimal recommendations that maximise the diversity of content users are exposed to, while still accounting for their preferences. Our method relies on a theoretical model that can sufficiently describe how content flows through the platform. We show that the model provides good approximations of empirical measures and demonstrate the effectiveness of the optimisation algorithm at mitigating the echo chamber effect on this dataset, even with limited budget for recommendations.

International audience

Last years have seen a regain of interest for the use of stochastic block modeling (SBM) in recommender systems. These models are seen as a flexible alternative to tensor decomposition techniques that are able to handle labeled data. Recent works proposed to tackle discrete recommendation problems via SBMs by considering larger contexts as input data and by adding second order interactions between contexts' related elements. In this work, we show that these models are all special cases of a single global framework: the Serialized Interacting Mixed membership Stochastic Block Model (SIMSBM). It allows to model an arbitrarily large context as well as an arbitrarily high order of interactions. We demonstrate that SIMSBM generalizes several recent SBM-based baselines. Besides, we demonstrate that our formulation allows for an increased predictive power on six real-world datasets., Comment: Published at ICDM 2022

International audience; Bitcoin is a cryptocurrency that stores transaction records in a public distributed ledger called the blockchain. All transactions that occurred since the beginning of Bitcoin in 2009 can therefore be consulted by anyone. This unique dataset allows us to study financial transaction networks among pseudonymous participants. Several works analyzed static transaction networks but did not consider the flow of money over the time. In this work, we focus on the analysis of flows, a challenging task given the scale of the data (hundreds of millions of transactions).We propose a method based on taint analysis to track Bitcoin money flow from initial starting points to the dissolution of the taint. The algorithm derives the dynamics subgraphs passing through known entities in the transaction network. We study the pattern of money flowing from different starting points: we taint coins minted by different mining pools in one day period between 2013 and 2016, and use graph embeddings from three representations of the data: (1) static network, (2) dynamic network, and (3) money flow pattern tree. Both qualitative and quantitative analysis show that mining pools have different diffusion patterns and that those patterns evolve over time. Based on this initial result, we are developing a method to select critical entities and expand our unsupervised approach to characterize other money flow patterns, in particular, related to illegal and cybercrime activities.

International audience; Finding a set of users that can maximize the spread of information in a social network is an important problem in social media analysis-being a critical part of several realworld applications such as viral marketing, political advertising and epidemiology. Although influence maximization has been studied extensively in the past, the majority of works focus on the algorithmic aspect of the problem, overlooking several practical improvements that can be derived by data-driven observations or the inclusion of machine learning. The main challenges of realistic influence maximization is on the one hand the computational demand of the diffusion models' repetitive simulations, and on the other the accuracy of the estimated influence spread. In this work, we propose CELFIE, an influence maximization method that utilizes learnt influence representations from diffusion cascades to overcome the use of diffusion models. It comprises of two parts. The first is based on INF2VEC, an unsupervised learning model that embeds influence relationships between nodes from a set of diffusion cascades. We create a new version of the model, based on observations from influence analysis on a large scale dataset, to match the scalability needs and the purpose of influence maximization. The second part capitalizes on the learned representations to redefine the traditional live-edge model sampling for the computation of the marginal gain. For evaluation, we apply our method in the Sina Weibo and Microsoft Academic Graph datasets, two large scale networks accompanied by diffusion cascades. We observe that our algorithm outperforms various baseline methods in terms of seed set quality and speed. In addition, the proposed INF2VEC modification for influence maximization provides substantial computational advantages in the price of a minuscule loss in the influence spread.

Graph autoencoders (GAE) and variational graph autoencoders (VGAE) emerged as two powerful groups of unsupervised node embedding methods, with various applications to graph-based machine learning problems such as link prediction and community detection. Nonetheless, at the beginning of this Ph.D. project, GAE and VGAE models were also suffering from key limitations, preventing them from being adopted in the industry. In this thesis, we present several contributions to improve these models, with the general aim of facilitating their use to address industrial-level problems involving graph representations.Firstly, we propose two strategies to overcome the scalability issues of previous GAE and VGAE models, permitting to effectively train these models on large graphs with millions of nodes and edges. These strategies leverage graph degeneracy and stochastic subgraph decoding techniques, respectively. Besides, we introduce Gravity-Inspired GAE and VGAE, providing the first extensions of these models for directed graphs, that are ubiquitous in industrial applications. We also consider extensions of GAE and VGAE models for dynamic graphs. Furthermore, we argue that GAE and VGAE models are often unnecessarily complex, and we propose to simplify them by leveraging linear encoders. Lastly, we introduce Modularity-Aware GAE and VGAE to improve community detection on graphs, while jointly preserving good performances on link prediction.In the last part of this thesis, we evaluate our methods on several graphsextracted from the music streaming service Deezer. We put the emphasis on graph-based music recommendation problems. In particular, we show that our methods can improve the detection of communities of similar musical items to recommend to users, that they can effectively rank similar artists in a cold start setting, and that they permit modeling the music genre perception across cultures. At the end of this thesis, we present two additional models, recently deployed in production on the Deezer service to recommend music to millions of users. While being less directly linked to GAE and VGAE models, they provide a complementary perspective on music recommendation topics related to the ones we previously studied., Les autoencodeurs de graphes (GAE) et les autoencodeurs variationnels de graphes (VGAE) se sont imposés comme deux puissants groupes de méthodes permettant de construire des représentations vectorielles des nœuds d'un graphe de manière non-supervisée, avec des applications à divers problèmes d'apprentissage tels que la prédiction de liens manquants et la détection de communautés de nœuds. Néanmoins, au début de ce projet de thèse, les GAE et VGAE souffraient de limitations majeures. Ces dernières entravaient l'utilisation de ces modèles dans le cadre d'applications industrielles. Dans cette thèse, nous présentons plusieurs contributions permettant d'améliorer les GAE et VGAE afin de faciliter de telles utilisations.Tout d'abord, nous proposons deux stratégies permettant de surmonter les problèmes de passage à l'échelle des GAE et VGAE, et d'entraîner ces modèles sur des graphes ayant des millions de nœuds et d'arêtes. Ces stratégies exploitent respectivement des techniques de dégénérescence de graphes et de décodage stochastique de sous-graphes. Par ailleurs, nous présentons nos GAE et VGAE "inspirés de la gravité" (de l'anglais "Gravity-Inspired GAE and VGAE"), qui constituent les premières extensions de ces modèles destinées aux graphes dirigés, qui sont omniprésents dans les applications industrielles. Nous étudions également des extensions destinées aux graphes dynamiques. En outre, nous démontrons que les GAE et VGAE existants sont souvent inutilement complexes, et nous proposons donc de les simplifier en ayant recours à des encodeurs linéaires. Enfin, nous présentons nos GAE et VGAE "informés par la modularité'' (de l'anglais "Modularity-Aware GAE and VGAE"), qui permettent d'améliorer la détection de communautés de nœuds, tout en préservant de bonnes performances pour la prédiction de liens manquants.Dans la dernière partie de cette thèse, nous évaluons nos méthodes sur plusieurs graphes extraits du service de streaming musical Deezer. Nous nous concentrons sur des problèmes de recommandation musicale à partir de graphes. En particulier, nous montrons que nos méthodes permettent d'améliorer la détection de communautés d'entités musicales à recommander aux mêmes utilisateurs, mais aussi de mieux classer des artistes similaires dans un contexte de "démarrage à froid", et enfin de mieux modéliser la perception des genres musicaux à travers différentes cultures. Pour terminer, nous présentons également deux autres modèles, récemment déployés en production chez Deezer afin de recommander de la musique à des millions d'utilisateurs. Bien qu'étant moins directement liés aux GAE et VGAE, ils fournissent un point de vue complémentaire sur des sujets de recommandation musicale connexes à ceux étudiés précédemment.

International audience

Online influence is the plinth of the social networks’ effect in our lives and its impact has been steeply increasing. From viral marketing to political campaigns and from news adoption to disease transmission, the way we are influenced by others is more prevalent than ever. In this thesis, we address the problem of efficiently learning and analyzing influence representations for numerous graph mining problems that are apropos.The first half of the thesis is devoted to the problem of influence maximization, an NP-hard combinatorial optimization problem. The aim is to find the nodes in a network that can maximize the spread of information, where the spread is typically defined by random influence probabilities and simple diffusion models. To address this, in the thesis' first part, we devise a node representation learning model based on diffusion cascades along with an adaptation of a traditional influence maximization algorithm that utilizes the output of the model. This framework surpasses competitive methods, evaluated in terms of computational time and the influence of the predicted seeds in cascades of the immediate future.The second part is devoted to learning how to perform influence maximization. We develop a graph neural network that inherently parameterizes an upper bound of influence estimation, and train it on small simulated graphs. We experimentally show that it can provide accurate estimations faster than the alternatives for graphs 10 times larger than the train set. Furthermore, we use the models’ predictions and representations to propose three new influence maximization methods. An adaptation of Cost Effective Lazy Forward that surpasses SOTA but with significant computational overhead, a Q-learning model that learns to retrieve seeds sequentially, and a submodular function that acts as proxy for the marginal gain and can be optimized adaptively and greedily with a theoretical guarantee. The latter strikes the best balance between efficiency and accuracy in our experiments.In the second half of the thesis, we focused on specific applications of influence in real data. In the third part we approach epidemic forecasting using influence learning. We utilize the inherent message passing of Graph Neural Networks to learn node representations based on mobility networks of a country’s regions and the history of the disease progression. These representations aim to capture how the epidemic diffuses through regions, and are used to predict the number of new COVID-19 cases with a forecasting window of up to 14 days. Furthermore, to capitalize on the lag of the COVID-19 spreading between countries, a meta-learning algorithm is proposed to transfer knowledge between models trained in some countries’ whole epidemic circle, to a model predicting cases for another country at the start of the outbreak, where the available training data is limited. Our approach outperforms baseline, time-series, and other deep learning models.In the final part, we analyze different versions of academic influence and device methods to quantify and predict it. Initially we utilize the Microsoft Academic Graph to build an author-citation network with billions of edges. We subsample it and perform directed-core decomposition to quantify it and visualize it through an interactive web-app. Subsequently we experiment with classifying the h-index of an author based on a GNN on her coauthorship graph and the text of her papers.We conclude the thesis with future directions regarding learning-based influence maximization with heterogeneous data and efficient neural network training through submodular active learning., L'influence en ligne est le socle de l'effet des réseaux sociaux sur nos vies et son impact n'a cessé de croître. Du marketing viral aux campagnes politiques à la transmission de maladies, la façon dont nous sommes influencés par les autres est plus répandue que jamais. Dans cette thèse, nous abordons le problème de l'apprentissage et de l'analyse efficaces des représentations d'influence pour de nombreux problèmes d'exploration de graphes qui sont à propos.La première moitié de la thèse est consacrée au problème de la maximisation de l'influence, un problème NP d'optimisation combinatoire. L'objectif est de trouver les nœuds d'un réseau qui peuvent maximiser la propagation de l'information, où la propagation est typiquement définie par des probabilités d'influence aléatoires et des modèles de diffusion simples. Pour répondre à ce problème, nous concevons un modèle d'apprentissage de la représentation des nœuds basé sur des cascades de diffusion ainsi qu'une adaptation d'un algorithme traditionnel de maximisation de l'influence qui utilise la sortie du modèle. Ce cadre surpasse les méthodes concurrentes, évaluées en termes de temps de calcul et d'influence des graines prédites dans les cascades du futur immédiat.La prochaine partie est consacrée à l'apprentissage de la maximisation de l'influence. Nous développons un réseau neuronal de graphe qui paramètre de manière inhérente une limite supérieure d'estimation de l'influence, et nous l'entraînons sur de petits graphes simulés. Nous montrons expérimentalement qu'il peut fournir des estimations précises plus rapidement que les autres solutions pour des graphes /dix/ fois plus grands que l'ensemble d'entraînement. En outre, nous utilisons les prédictions et les représentations des modèles pour proposer nouvelles méthodes de maximisation de l'influence. Une adaptation de un ancien algorithme, un modèle de Q-learning, et une fonction submodulaire qui agit comme un proxy pour le gain marginal et peut être optimisée de manière adaptative et avide avec certaines garanties théoriques. Cette dernière fonction offre le meilleur équilibre entre efficacité et précision.Dans la deuxième moitié de la thèse, nous sommes concentrés sur des applications spécifiques de l'influence. Nous abordons la prévision des épidémies en utilisant l'apprentissage par influence. Nous utilisons le passage de messages inhérent aux réseaux neuronaux graphiques pour apprendre des représentations de nœuds basées sur les réseaux de mobilité des régions d'un pays et l'histoire de la contagion. Ces représentations sont utilisées pour prédire le nombre de nouveaux cas de COVID-19 avec une fenêtre de prévision allant jusqu'à 14 jours. En outre, pour tirer parti du décalage de la propagation entre les pays, un algorithme de méta-learning est proposé pour transférer les connaissances entre les modèles formés dans le cercle épidémique complet de certains pays, à un modèle prédisant les cas pour un autre pays au début de l'épidémie, où les données d'apprentissage disponibles sont limitées. Notre approche surpasse les modèles de référence, les séries temporelles et d'autres modèles d'apprentissage profond.Dans la dernière partie, nous analysons différentes versions de l'influence académique et essayons de la prédire. Dans un premier temps, nous utilisons le MAG pour construire un réseau de citations d'auteurs avec des milliards d'arêtes. Nous le sous-échantillonnons et effectuons une décomposition en noyaux dirigés pour le quantifier et le visualiser au moyen d'une application web interactive. Ensuite, nous expérimentons la classification du h-index d'un auteur sur la base d'un GNN sur son graphe de coauteurs et le texte de ses articles. Nous concluons la thèse en abordant de futures directions concernant la maximisation de l'influence basée sur l'apprentissage avec des données hétérogènes et l'entraînement efficace des réseaux neuronaux par l'apprentissage actif submodulaire.

Les étudiants en apprentissage ont été fortement impactés par les confinements successifs de la Covid-19, nous avons pu les observer durant trois ans pour analyser leur capacité à travailler collectivement avec la plateforme numérique collaborative MS Teams. C’est dans le contexte universitaire que nous contribuons à la connaissance des dynamiques des groupes de travail qui utilisent un système d’information pour communiquer, partager des documents en mobilité et coconstruire. Grâce au modèle de l’activité collective (Engeström, 2001) nous analysons les dynamiques des équipes dans les salles de coworking et dans les espaces numériques. Les entretiens des étudiants et formateurs sont croisés avec les réponses aux questionnaires individuels et les indicateurs numériques d’activité de la plateforme en ligne. Nos résultats mettent en évidence trois organisations distinctes du travail : 1) Individuel : travaux réalisés seuls ou en binôme avec peu d’échange, 2) Coopératif : travaux réalisés collectivement avec une division forte des tâches organisée par un leader et 3) Collaboratif : travaux coconstruits avec de nombreux échanges entre tous les membres. Les traces numériques dans les espaces MS Teams participent à l’identification des phases qui se succèdent durant les projets. Nous constatons que l’usage des outils numériques collaboratifs influence l’organisation des petits groupes de travail, nos résultats aident à comprendre les activités collectives dans les SI afin d’en améliorer le management.

Laws and administrative rules have been addressing disability and accessibility, through quotas and financial penalties, or students social criteria assessment on French national ranking platforms. Such official measures enable minorities and other discriminated groups to be represented.As half of the humanity does not constitute a minority, no quota policy is supposed to be applied, leaving gender imbalance as a potential issue. Academics have launched various inquiries and studies on this issue (Mothers in Science, UNESCO I’d Blush if I Could, etc.), that is crucial because of the female under-representation in scientific fields in STEM (Science, Technology, Engineering, and Mathematics).“Gender equality paradox”names the observation that the under-representation of women in scientific fields (in particular, those related to STEM) is stronger in the most developed countries. As one says: “We are open. The door is just very heavy”. The door is not closed but women must get a foot in the door to be invited, integrated, accepted, and hired. Women fight to break the glass ceiling in many professions. At the same time, as they can encounter obstacles, the sticky floor restricts them to relative non-strategic positions and prevents their scientific career from really taking off.The tooth paste tube phenomenon is also invoked as a metaphor that illustrates how few female colleagues must assume the committees and representations for the community to comply with F/M quotas when required for hiring or mediating, for instance.The consequence can be an over-solicitation and over-impact on the communities. The facts are horrendous, and actions have hence to be taken to reach a global balance with diversity, inclusiveness and fair cohabitation without discrimination. All the launched initiatives testify this awareness. As a warning underlies the need to support these actions, one must take care of a negative effect on under-represented members by (unfairly) reinforcing their impostor syndrome.Inside the science itself, and not only in terms of population or scientometrics, but a new phenomenon also arises with the evolution of topics, technics, and tools. One striking example is the risk with gender and minority “invisibilization”in AI, as the social representation biases are emphasized (few entries in Wikipedia, Matilda effect, lack of historical figures and illustration, rare “role models”in science,...) with unbalanced learning data. The dual issue is how rule-based systems and by whom, if any bias here also, are encoded to be aware of all the diversity any decision implies.

As social networks are ubiquitous in everyday life, problems such as misinformation, bots and polarisation of opinion gain more and more importance. This paper focuses on the last one as we propose novel methods to reduce the amount of polarisation in a social group. We leverage the voter model in which users hold binary opinions and repeatedly update their beliefs based on others they connect with. Stubborn agents who never change their minds ("zealots") are also disseminated through the network. We are interested in two equilibrium metrics, the diversity of opinions $\sigma$ and the density of active links $\rho$. The latter is a measure of the exposure to adverse opinions. We derive formulas to compute them that are valid in any directed, weighted network. Then we study the problem of finding optimal numbers of zealots in order to maximise these quantities. We account for the presence of a backfire effect, which may lead the group to react negatively and reinforce its level of polarisation in response. We provide exact solutions in the specific context of a complete unweighted graph, and propose a method for $\sigma$ in the general case. Finally we apply these problems to the network of the US House of Representatives. The network exhibits high levels of opinion diversity but lower levels of active links density. We find optimal numbers of zealots that maximise these quantities and show that both can be efficiently increased in some cases.

As social networks take an ever-prominent role in information access, combating disinformation becomes increasingly important. Given current volumes of data, automated approaches for the detection of disinformation are the only ones able to offer an online solution. We formalize disinformation as an anomaly in the organic evolution of the network, in terms of users, content or coordination. To identify it, we propose a solution based on Temporal Graph Networks (TGN) adapted to the detection of anomalies and completed with a reliability module which permits users to trade precision for recall. Inheriting the performances of TGN, this solution is able to scale up, work on continuous time settings, and handle multimodality (text, image and video). Moreover, when compared to existing models, our approach outperforms state-of-the-art solutions for anomaly detection in dynamic graphs on several classical datasets. Lastly, it has been tested on a Twitter dataset of the French presidential election of 2022, providing useful insights on manipulation during the campaign.

Since the development of writing 5000 years ago, human-generated data gets produced at an ever-increasing pace. This rate has been greatly influenced by technical innovations, such as clay tablets, papyrus, paper, press, and more recently the Internet. At the same time, new methods designed to handle and archive these growing information flows emerged: clay archives (Nippur, Mari), early libraries (Alexandria, Rome's Tabularia, Athens' Metroon), religious scriptoriums (abbeys, monasteries), modern libraries and, more recently, machine learning. Each of these archival methods aims at easing information retrieval.Nowadays, archiving is not enough anymore. The amount of data that gets generated daily is beyond human comprehension, and appeals for new information retrieval strategies.Instead of referencing every single data piece as in traditional archival techniques, a more relevant approach consists in understanding the overall ideas conveyed in data flows. To spot such general tendencies, a precise comprehension of the underlying data generation mechanisms is required.In the rich literature tackling this problem, the question of information interaction remains nearly unexplored. Explicitly, few works explored the influence of anterior human-generated data on ulterior data creation mechanisms. In this manuscript, we develop a panel of new machine learning methods that explore this specific aspect of online data generation.First, we investigate the frequency of such interactions. Building on recent advances made in Stochastic Block Modelling, we explore the role of interactions in several social networks. We find that interactions are rare in these datasets. Then, we wonder how interactions evolve over time. Earlier data pieces should not have an everlasting influence on ulterior data generation mechanisms; an ad may exert a short-term influence on buying behaviours, but would have no influence on them a decade later for instance. We model this using dynamic network inference advances on social media datasets. We conclude that interactions are brief and that their intensity typically decays in an exponential fashion.Finally, as an answer to the previous points, we design a framework that jointly models rare and brief interactions. Doing so, we exploit a recent bridge between Dirichlet processes and Point processes. We improve on this advance and discuss the more general Dirichlet-Point processes. We argue that this new class of models readily fits brief and sparse interaction modelling. We conduct a large-scale application on Reddit and find that interactions play a minor role in this dataset. From a broader perspective, our work results in a collection of highly flexible models and in a rethinking of core concepts of machine learning. Consequently, we open a range of novel perspectives both in terms of real-world applications and in terms of technical contributions to machine learning., À compter du développement de l'écriture, il y a environ 5000 ans, la quantité de données générées par l'homme n'a cessé de croître de façon exponentielle. Cette tendance est étroitement liée au rythme des innovations techniques facilitant leur diffusion ; chacune de ces avancées va de pair avec le développement de méthodes visant à faciliter leur archivage. Or, aujourd’hui, l'archivage classique ne peut répondre à nos besoins. L’immense quantité de données générée quotidiennement requiert de nouvelles stratégies de récupération de l'information. Plutôt que référencer individuellement chaque fragment d'information (un tweet, un mème, une vidéo), comme le ferait une méthode d'archivage standard, une approche plus pertinente consiste à comprendre les principales idées convoyées par les flux d'informations à une échelle globale. Discerner de telles tendances à partir d'une masse de données apparemment informe nécessite une compréhension fine des mécanismes de génération de données sous-jacentsDans la riche littérature traitant du fonctionnement de tels mécanismes, la question de l'interaction entre différents fragments d'information est restée largement inexplorée. Plus explicitement, peu de travaux ont étudié le rôle de l'influence de données produites antérieurement par une population sur les mécanismes génératifs de données produites ultérieurement par cette même population. Dans ce manuscrit, nous développons un éventail de nouvelles méthodes d'apprentissage automatique, permettant d'explorer cet aspect particulier des mécanismes de génération de données en ligne.En premier lieu, nous examinons la fréquence de telles interactions entres fragments d'information. Sur la base de récentes avancées en Stochastic Block Modelling, nous explorons le rôle de ces interactions dans la génération d'informations sur plusieurs plateformes sociales en ligne. Cette première étude nous permet de conclure que les interactions sont rares dans ces jeux de données.Ensuite, nous nous intéressons à l'évolution temporelle de ces interactions ; un fragment d'information ne devrait pas influencer indéfiniment les mécanismes de génération ultérieurs, ni de la même manière au cours du temps. Une publicité peut par exemple influer les comportements d'achat à court terme, mais pas nécessairement à l'échelle d'une décennie. Nous modélisons ce phénomène dans des jeux de données provenant des réseaux sociaux, en nous basant sur de récents progrès effectués en inférence de graphes dynamiques. Nous en tirons la conclusion que les interactions sont brèves, et que leur intensité décroît généralement de manière exponentielle avec le temps.Enfin, en guise de réponse aux difficultés soulevées dans les deux paragraphes précédents, nous élaborons une méthode permettant de modéliser des interactions rares et brèves. Ce faisant, nous étudions des connexions encore peu exploitées entre les processus de Dirichlet et les processus ponctuels. Nous explorons cette voie en vue de la résolution de notre problème, et discutons à cette occasion le concept plus général de processus de Dirichlet-Point. Nous justifions l'usage de cette nouvelle classe de modèles dans la modélisation d'interactions rares et brèves. En guise de conclusion, nous conduisons une application à grande échelle de notre méthode sur un jeu de données provenant de Reddit ; suite à cette étude, nous concluons que les interactions jouent un rôle mineur dans ce cas précis.D'un point de vue plus global, nos travaux aboutissent à une collection de modèles flexibles en termes de champs d'applications, d'une part, et à une réflexion sur plusieurs concepts au coeur des méthodes d'apprentissage automatique actuelles d'autre part. Ce faisant, nous ouvrons la voie à une panoplie de nouvelles perspectives, tant en termes d'applications à des problématiques de recherche transdisciplinaires, qu'en termes de contributions à l'apprentissage automatique en tant que tel.

This paper introduces G-Impact, an agent-based model that combines modelling of household consumption and belief diffusion. Household decisions integrate personal impacts (quality, cost), perceived consequences (climate change, human responsibility), and social norms. The evaluation of these different criteria relies on household beliefs, which can be exchanged during social interactions. These beliefs can be used to explain household decisions on a macro and micro scale, and thus to target information or incentive policies. The model is applied to dietary choice in France, among the omnivorous diet (INCA3), the flexitarian diet and the vegetarian diet. The costs and greenhouse gases emissions of the different diets are initialized from real data. The initial beliefs of households are derived from opinion surveys. In the control simulation, we observe a significant increase in the proportion of flexitarians, and a slight increase in the proportion of vegetarians. We also illustrate the need to properly inform households with the emergence of fake news.

TSN (Time Sensitive Networking) standards are arousing growing interest in the critical on-board network community because of their promise of deterministic Ethernet networking. Among these standards IEEE802.1AS allows the synchronization of network devices. This protocol achieves much greater precision than other Ethernet synchronization standard such as NTP or PTP. Thus, it paves the way for new network mechanisms, such as the Time Aware Shaper, and allows the use of applications that are more constrained in terms of synchronization. In order to study the new mechanisms allowing to reach this precision, precise simulations are mandatory. In this paper, we review the different existing tools, extend the most promising one to incorporate the most advanced features of IEEE802.1AS and assess its behavior with respect to measurements. More specifically, we extend an OMNeT++ simulation library, calibrate its results following an extensive measurement campaign of switched Ethernet devices supporting IEEE 802.1AS to assess its performance in terms of precision., Les normes TSN (Time Sensitive Networking) suscitent un intérêt croissant dans la communauté des réseaux critiques embarqués en raison de leur promesse d'un réseau Ethernet déterministe. Parmi ces normes, IEEE802.1AS permet la synchronisation des dispositifs du réseau. Ce protocole atteint une précision bien supérieure à celle des autres normes de synchronisation Ethernet telles que NTP ou PTP. Ainsi, il ouvre la voie à de nouveaux mécanismes de réseau, tels que le Time Aware Shaper, et permet l'utilisation d'applications plus contraintes en termes de synchronisation. Afin d'étudier les nouveaux mécanismes permettant d'atteindre cette précision, des simulations précises sont obligatoires. Dans cet article, nous passons en revue les différents outils existants, nous étendons le plus prometteur d'entre eux afin d'intégrer les fonctionnalités les plus avancées de IEEE802.1AS et nous évaluons son comportement par rapport aux mesures. Plus précisément, nous étendons une bibliothèque de simulation OMNeT++, calibrons ses résultats suite à une campagne de mesure extensive de dispositifs Ethernet commutés supportant IEEE 802.1AS afin d'évaluer ses performances en termes de précision.

Detection and characterization of polarization are of major interest in Social Network Analysis, especially to identify conflictual topics that animate the interactions between users. As gatekeepers of their community, users in the boundaries significantly contribute to its polarization. We propose ERIS, a formal graph approach relying on community boundaries and users' interactions to compute two metrics: the community antagonism and the porosity of boundaries. These values assess the degree of opposition between communities and their aversion to external exposure, allowing an understanding of the overall polarization through the behaviors of the different communities. We also present an implementation based on matrix computations, freely available online. Our experiments show a significant improvement in terms of efficiency in comparison to existing solutions. Finally, we apply our proposal on real data harvested from Twitter with a case study about the vaccines and the COVID-19.

Among the issues the information system security community has to fix, the security of both data and algorithms is a concern. The security of algorithms is dependent on the reliability of the input data. This reliability is questioned, especially when the data is generated by humans (or bots operated by humans), such as in online social networks. Event detection algorithms are an example of technology using this type of data, but the question of the security is not systematically considered in this literature. We propose in this paper a first contribution to a threat model to overcome this problem. This threat model is composed of a description of the subject we are modelling, assumptions made, potential threats and defence strategies. This threat model includes an attack classification and defensive strategies which can be useful for anyone who wants to create a resilient event detection algorithm using online social networks.

Automatic detection of relevant groups of nodes in large real-world graphs, i.e. community detection, has applications in many fields and has received a lot of attention in the last twenty years. The most popular method designed to find overlapping communities (where a node can belong to several communities) is perhaps the clique percolation method (CPM). This method formalizes the notion of community as a maximal union of $k$-cliques that can be reached from each other through a series of adjacent $k$-cliques, where two cliques are adjacent if and only if they overlap on $k-1$ nodes. Despite much effort CPM has not been scalable to large graphs for medium values of $k$. Recent work has shown that it is possible to efficiently list all $k$-cliques in very large real-world graphs for medium values of $k$. We build on top of this work and scale up CPM. In cases where this first algorithm faces memory limitations, we propose another algorithm, CPMZ, that provides a solution close to the exact one, using more time but less memory., Comment: 15 pages, 5 figures, 1 table

International audience

The textual content of a document and its publication date are intertwined. For example, the publication of a news article on a topic is influenced by previous publications on similar issues, according to underlying temporal dynamics. However, it can be challenging to retrieve meaningful information when textual information conveys little. Furthermore, the textual content of a document is not always correlated to its temporal dynamics. We develop a method to create clusters of textual documents according to both their content and publication time, the Powered Dirichlet-Hawkes process (PDHP). PDHP yields significantly better results than state-of-the-art models when temporal information or textual content is weakly informative. PDHP also alleviates the hypothesis that textual content and temporal dynamics are perfectly correlated. We demonstrate that PDHP generalizes previous work --such as DHP and UP. Finally, we illustrate a possible application using a real-world dataset from Reddit., Comment: in French

Although 5G is designed to provide a large range of new digital services such as immersive experiences, connected machines, and Internet of Things (IoT), 5G networks may not be able to completely meet the needs of emerging novel applications, as well as to address the future needs of the business and regulatory ecosystem. Beyond 5G (B5G) and Sixth Generation (6G) wireless systems are expected to overcome 5G network limits and support a broader range of new use cases in this rapidly expanding digital era. In this survey article, we provide a summary of the major 5G networks shortcomings that are discussed in the literature and classify them into a taxonomy of six primary domains. For each of these domains, we also highlight and discuss the research challenges that will be faced in the development of future B5G and 6G networks.

We try to build a musical taste fingerprint based on declared preferences of users from a music streaming service. We propose two distinct methods, that turn out to be conflicting : one that captures the uniqueness,and one that pictures the very essence of one's preferences.

This document describes the constitution of the FOPPA database (French Open Public Procurement Award notices), which relies on a subset of the TED database that contains French public procurement notices published from 2010 to 2020. These data have a number of issues, the most serious being that the unique IDs of most involved agents are missing. We explain how we solve these issues to produce a usable database. This work was conducted under the framework of the ANR DeCoMaP project (ANR-19-CE38-0004).

Graph Neural Networks (GNNs) have achieved great successes in many learning tasks performed on graph structures. Nonetheless, to propagate information GNNs rely on a message passing scheme which can become prohibitively expensive when working with industrial-scale graphs. Inspired by the PPRGo model, we propose the CorePPR model, a scalable solution that utilises a learnable convex combination of the approximate personalised PageRank and the CoreRank to diffuse multi-hop neighbourhood information in GNNs. Additionally, we incorporate a dynamic mechanism to select the most influential neighbours for a particular node which reduces training time while preserving the performance of the model. Overall, we demonstrate that CorePPR outperforms PPRGo, particularly on large graphs where selecting the most influential nodes is particularly relevant for scalability. Our code is publicly available at: https://github.com/arielramos97/CorePPR., Comment: Accepted at the "NeurIPS 2022 New Frontiers in Graph Learning Workshop (NeurIPS GLFrontiers 2022)"

Social media is often described as both reflecting and distorting real-life debates. Indeed, social division occurs not only offline but also online on various political topics or scientific controversies. Several studies propose tools to identify and quantify online controversies through stance detection or polarization measures. While polarization is typically studied as a "snapshot" in time of a social network, we consider it as a temporal process. Moreover, the recent evolution in temporal graph representation learning provides new tools to directly combine time, content and graph topology. Current techniques that characterize polarization are beginning to use these tools, typically using a Lagrangian description which focuses on user trajectories. In this article, we make a case for approaching these problems with a Eulerian description, the concurrent description in fluid mechanics. In this description, the temporal evolution of nodes embeddings is represented with a deformation of velocity vector fields. Finally, we validate our method on a retweet graph from the last French presidential election campaign.

In order to study real-world systems many works model them through signed graphs, i.e. graphs whose edges are labeled as either positive or negative. Such a graph is considered as structurally balanced when it can be partitioned into a number of modules, such that positive (negative) edges are located inside (in-between) the modules. When it is not the case, authors look for the closest partition to such balance, a problem called Correlation Clustering (CC). The standard approach used in the literature is to find a single partition and focus the rest of the analysis on it, as if it was sufficient to fully characterize the studied system. Yet, it may not reflect the structure of the network, and one may need to seek for other partitions to build a better picture.

National audience; Our objective is to detect abusive messages in online conversations through content- and graph-based methods. Our approach consists in: 1) Exhibiting the importance of structural information in the context of online abuse detection; 2) Combining textual and structural information by proposing fusion methods of both types of information.

Les réseaux sociaux peuvent être un support opérationnel pour la détection d'évènements. La littérature se concentre sur l'amélioration de la performance des algorithmes de détection d'évènements, mais fait peu mention de mesures permettant de se défendre contre une potentielle manipulation de la donnée d'entrée (appelé "adversarial learning" ou encore "apprentissage adverse"). Toute personne le souhaitant peut poster des messages sur les réseaux sociaux, il est donc possible pour n'importe qui de manipuler les données utilisées pour la détection. En conséquence, les algorithmes de détection d'évènements actuels sont particulièrement vulnérables à l'apprentissage adverse. Cet article propose une première typologie pour décrire les menaces liées à l'apprentissage adverse contre les algorithmes de détection d'évènements.

Most models of information diffusion online rely on the assumption that pieces of information spread independently from each other. However, several works pointed out the necessity of investigating the role of interactions in real-world processes, and highlighted possible difficulties in doing so: interactions are sparse and brief. As an answer, recent advances developed models to account for interactions in underlying publication dynamics. In this article, we propose to extend and apply one such model to determine whether interactions between news headlines on Reddit play a significant role in their underlying publication mechanisms. After conducting an in-depth case study on 100,000 news headline from 2019, we retrieve state-of-the-art conclusions about interactions and conclude that they play a minor role in this dataset., Comment: Published at the conference Complex Networks and their Applications

International audience

One of the most common approaches to the analysis of dynamic networks is through time-window aggregation. The resulting representation is a sequence of static networks, i.e. the snapshot graph. Despite this representation being widely used in the literature, a general framework to evaluate the soundness of snapshot graphs is still missing. In this article, we propose two scores to quantify conflicting objectives: Stability measures how much stable the sequence of snapshots is, while Fidelity measures the loss of information compared to the original data. We also develop a technique of targeted filtering of the links, to simplify the original temporal network. Our framework is tested on datasets of proximity and face-to-face interactions., Comment: 12 pages, 5 figures. To be published in "Proceedings of the Tenth International Conference on Complex Networks and Their Applications" Replacement edits: corrected a misspelled address, improved image quality (the content was not modified)

International audience

International audience

The amount and variety of data is increasing drastically for several years. These data are often represented as networks, which are then explored with approaches arising from network theory. Recent years have witnessed the extension of network exploration methods to leverage more complex and richer network frameworks. Random walks, for instance, have been extended to explore multilayer networks. However, current random walk approaches are limited in the combination and heterogeneity of network layers they can handle. New analytical and numerical random walk methods are needed to cope with the increasing diversity and complexity of multilayer networks. We propose here MultiXrank, a Python package that enables Random Walk with Restart (RWR) on any kind of multilayer network with an optimized implementation. This package is supported by a universal mathematical formulation of the RWR. We evaluated MultiXrank with leave-one-out cross-validation and link prediction, and introduced protocols to measure the impact of the addition or removal of multilayer network data on prediction performances. We further measured the sensitivity of MultiXrank to input parameters by in-depth exploration of the parameter space. Finally, we illustrate the versatility of MultiXrank with different use-cases of unsupervised node prioritization and supervised classification in the context of human genetic diseases.

International audience; We address formally the problem of opinion dynamics when the social network composed of conformists and contrarians are not only influenced by their neighbors, but also by an external influential entity referred to as a marketer. The population of contrarians tries to have an opinion that is the opposite of the opinion held by the conformists. The influential entity tries to sway the overall opinion as close as possible to the desired opinion by using a specific influence budget. The main technical issue addressed is finding how the marketer should allocate its budget among the agents such that the agents' opinion will be as close as possible to the desired opinion while taking into account the behavior of the contrarian population. Our main results show that the marketer has to prioritize certain agents over others based on their initial condition, their influence power in the social graph and the population class they belong to. Numerical examples illustrate the analysis.

International audience; We study in this paper crowdfunding from the platform perspective. We consider a platform where several project campaigns are running and competing on fund collection. The platform manager's objective is to increase the proportion of successful projects and he uses the lever of project promotion on the platform website to influence the funders' and balance the potential future funds between projects. We first consider the platform policy derivation as an optimization problem and show to derive the optimal project promotion policy. We then consider the more realistic setting where the platform manager adapts his policy online when observing the status of fund collection. We model this online strategy as a real option and develop a dynamic programming algorithm that finds the optimal strategy depending on the observed project campaign status.

International audience; Identifying influential nodes in social networks is a fundamental issue. Indeed, it has many applications, such as inhibiting epidemic spreading, accelerating information diffusion, preventing terrorist attacks, and much more. Classically, centrality measures quantify the node's importance based on various topological properties of the network, such as Degree and Betweenness. Nonetheless, these measures are agnostic of the community structure, although it is a ubiquitous characteristic encountered in many real-world networks. To overcome this drawback, there is a growing trend to design so-called community-aware centrality measures. Although several works investigate the relationship between various classical centrality measures [1-3], the interplay between classical and community-aware centrality measures is still unexplored. This work presents an extensive investigation aimed at a better understanding of the relationship between community-aware centrality measures, classical centrality measures, and network topology. Artificial and real-world networks are used in the experiments. The Kendall's Tau correlation quantifies the interaction between ten classical and twenty-eight community-aware centrality measures. The community-aware centrality measures are divided into three groups. The first group's ten measures are based on the intra-community links of a node (local measures). The second group's twelve measures are based on the inter-community links of a node (global measures). Finally, the six measures of the third group consider both types of links (mixed measures). The LFR algorithm generates artificial networks with controlled properties. Indeed, the community structure strength (µ), the exponent of the degree distribution (γ), and the community size distribution (θ) can be specified. The experiments show that the community structure strength is the main feature governing the correlation between classical and community-aware centrality measures. The heatmap on the left of Figure 1 represents the correlation in an artificial network with a strong community structure. The global community-aware centrality measures exhibit a low correlation with classical centrality measures, while local communityaware centrality measures show a high correlation. The results are inverted when the network has a weak community structure. Differences are more subtle for mixed community-aware measures. One can also notice that results are relatively insensitive to variation of the degree and community size distributions' exponents. Fifty real-world networks originating from various domains are also investigated. Linear regression is performed considering six macroscopic (

Social networks (SN) are omnipresent in our lives today. Not all users have the same behavior on these networks. If some have a low activity, rarely posting messages and following few users, some others at the other extreme have a significant activity, with many followers and regularly posts. The important role of these popular SN users makes them the target of many applications for example for content monitoring or advertising. After a study of the metadata of these users, in order to detect abnormal accounts, we present an approach allowing to detect users who are becoming popular. Our approach is based on modeling the evolution of popularity in the form of frequent patterns. These patterns describe the behaviors of gaining popularity. We propose a pattern matching model which can be used with a data stream and we show its scalability and its performance by comparing it to classic models. Finally, we present a clustering approach based on PageRank. This work allow to identify groups of users sharing the same role, using the interaction graphs, Les réseaux sociaux sont devenus des outils de communication primordiaux et sont utilisés quotidiennement par des centaines de millions d’utilisateurs. Tous ces utilisateurs n’ont pas le même comportement sur ces réseaux. Si certains ont une faible activité, publient rarement des messages et suivent peu d’utilisateurs, d’autres, à l’opposé, ont une activité importante, avec de nombreux abonnés et très publient régulièrement. Le rôle important de ces utilisateurs influents en font des cibles intéressantes pour de nombreuses applications, par exemple pour la surveillance ou la publicité. Après une étude des méta-donées de ces utilisateurs, afin de d´détecter des comptes anormaux, nous présentons une approche permettant de d´détecter des utilisateurs devenant populaires. Notre approche s’appuie sur une modélisation de l’évolution de la popularité sous la forme de motifs fréquents. Ces motifs décrivent les comportements de gain en popularité. Nous proposons un modèle de matching des motifs permettant d’être utilisé avec un flux de données et, nous montrons sa capacité à passer à l’échelle en le comparant à des modèles classiques. Enfin, nous présentons une approche de clustering basé sur le PageRank. Ces travaux permettent d’identifier des groupes d’utilisateurs partageant le même rolex, en utilisant les graphes ’interactions qu’ils génèrent.

International audience

The rapid emergence of social networks, their major impact on today's society, and the recent access to large amounts of data about them has led to a strong interest in the study of complex networks in last decades. Many random graph models have been proposed to reproduce these networks and their properties - small diameter, high clustering coefficient, power-law degree distribution, presence of communities, ... However, due to the complexity of their theoretical study, the proposed models are often designed for undirected graphs and focus on the emergence of one or two properties at a time. This thesis aims at developing models of random graphs that are general enough to reproduce many complex properties observed in real-world networks. In particular, we present models for constructing graphs with arbitrary degree distributions, directed graphs with a high clustering coefficient, and hypergraphs with power-law degree distributions and a strong presence of communities. We study analytically and experimentally the properties of the presented models. We develop a tool using Markov chains to compute the degree distribution of preferential attachment models.In order to help in the construction of these models, we study in this thesis two large complex networks: a directed network with all followings on Twitter, with 505 million accounts and 23 billion followings; and a hypergraph of scientific co-publications extracted from the Scopus database, containing 2.2 million authors and 3.9 million publications. Atypical properties emerge from the study of these two graphs, coming in particular from their directed and hypergraph character, that no model of the literature could reproduce until now., L'émergence rapide des réseaux sociaux lors des deux dernières décennies, leur impact majeur sur la société actuelle, ainsi que l'accès récent à de grandes quantités de données les concernant, a amené un fort attrait pour l'étude des réseaux complexes. De nombreux modèles de graphes aléatoires ont été proposé afin de reproduire ces réseaux et leurs propriétés - faible diamètre, distribution des degrés en loi de puissance, présence de communautés… Cependant, du fait de la complexité de l'étude théorique de leurs propriétés, les modèles proposés sont souvent conçus pour des graphes non orientés et se concentrent sur l'émergence d'une ou deux propriétés à la fois.Cette thèse a pour but de développer des modèles de graphes aléatoires suffisamment généraux pour reproduire de nombreuses propriétés complexes observées dans les réseaux du monde réel. En particulier, nous présentons des modèles permettant de construire des graphes avec des distributions de degrés quelconques, des graphes dirigés avec un haut coefficient de clustering, et des hypergraphes avec distribution de degrés en loi de puissance et une forte présence de communautés. Nous étudions analytiquement et expérimentalement les propriétés des modèles présentées. Nous développons enfin un outil utilisant les chaînes de Markov pour calculer la distribution des degrés de modèles d'attachement préférentiel.Afin de s'aider dans la construction de ces modèles, nous étudions dans cette thèse deux réseaux complexes de grande taille : un graphe dirigé de l'ensemble des liens d'abonnements sur le réseau social Twitter, avec 505 millions de comptes et 23 milliards d'abonnements ; et un hypergraphe de co-publications scientifiques extraites de la base de données Scopus, contenant 2.2 millions d'auteurs distincts et 3.9 millions de publications. Des propriétés atypiques émergent de l'étude de ces deux graphes, provenant en particulier de leur caractère dirigé et d'hypergraphe, et qu'aucun modèle de la littérature ne permettait jusqu'alors de reproduire.

International audience; Community detection in multilayer networks aims to identify groups of well-connected nodes across multiple layers. While existing methods have been developed to deal with large graphs with few layers (typically less than 10), many real-world datasets are structured by transitive relationships that give rise to networks with thousands of extremely dense layers (e.g. co-citation networks, IMDb actor graphs, co-reference information network, social networks). In addition, in these datasets the layers are often associated with textual summaries which provide important and hitherto unexploited information on the nature of the relation encoded by the layer. In this paper, we propose a new method which exploits the text associated with the layers in order to identify communities grouping together nodes connected through several semantically close layers. The method consists in embedding layer textual information in an Euclidean space, and to use it to group together, in the same community, nodes belonging to semantically close layers. To that end, we develop a pattern mining approach that extracts communities from numerical data. This approach, which mixes both symbolic and numeric techniques, is particularly well suited to identify communities in multilayer graphs. Indeed, we show that it obtains more diverse and better quality communities than those obtained by state-of-the-art competitors on datasets where ground truth is known. We also show that taking into account the semantic information improves the quality of the communities.

Preprints promote the open and fast communication of non-peer reviewed work. Once a preprint is published in a peer-reviewed venue, the preprint server updates its web page: a prominent hyperlink leading to the newly published work is added. Linking preprints to publications is of utmost importance as it provides readers with the latest version of a now certified work. Yet leading preprint servers fail to identify all existing preprint–publication links. This limitation calls for a more thorough approach to this critical information retrieval task: overlooking published evidence translates into partial and even inaccurate systematic reviews on health-related issues, for instance. We designed an algorithm leveraging the Crossref public and free source of bibliographic metadata to comb the literature for preprint–publication links. We tested it on a reference preprint set identified and curated for a living systematic review on interventions for preventing and treating COVID-19 performed by international collaboration: the COVID-NMA initiative (covid-nma.com). The reference set comprised 343 preprints, 121 of which appeared as a publication in a peer-reviewed journal. While the preprint servers identified 39.7% of the preprint–publication links, our linker identified 90.9% of the expected links with no clues taken from the preprint servers. The accuracy of the proposed linker is 91.5% on this reference set, with 90.9% sensitivity and 91.9% specificity. This is a 16.26% increase in accuracy compared to that of preprint servers. We release this software as supplementary material to foster its integration into preprint servers’ workflows and enhance a daily preprint–publication chase that is useful to all readers, including systematic reviewers. This preprint–publication linker currently provides day-to-day updates to the biomedical experts of the COVID-NMA initiative.