Search

In this work, we tackle the problem of intersectional group fairness in the classification setting, where the objective is to learn discrimination-free models in the presence of several intersecting sensitive groups. First, we illustrate various shortcomings of existing fairness measures commonly used to capture intersectional fairness. Then, we propose a new framework called the $\alpha$ Intersectional Fairness framework, which combines the absolute and the relative performances between sensitive groups. Finally, we provide various analyses of our proposed framework, including the min-max and efficiency analysis. Our experiments using the proposed framework show that several in-processing fairness approaches show no improvement over a simple unconstrained approach. Moreover, we show that these approaches minimize existing fairness measures by degrading the performance of the best of the group instead of improving the worst.

International audience; Training sequence models such as transformers with symbolic music requires a representation of music as sequences of atomic elements called tokens. State-of-the-art music tokenizations encode pitch values explicitly, which complicates the ability of a machine learning model to generalize musical knowledge at different keys. We propose tracks for a tokenization encoding pitch intervals rather than pitch values, resulting in transposition invariant representations. The musical expressiveness of this new tokenization is evaluated through two MIR classification tasks: composer classification and end of phrase detection. We release publicly the code produced in this research.

National audience; La discipline du Traitement Automatique du Langage Naturel (TALN) a connu d’importants progrès ces dix dernières années grâce aux avancées de l’intelligence artificielle et en particulier des réseaux de neurones profonds. Ces techniques sont aujourd’hui largement utilisées pour l’extraction de connaissances, l’analyse de sentiments ou encore la traduction automatique de textes. Les similarités structurelles et conceptuelles entre le langage naturel et la musique ont motivé de nombreuses initiatives de recherche visant à adapter les outils du TALN pour le traite- ment de données musicales symboliques ou audio. Ces démarches ont fourni des résultats prometteurs, notamment dans les domaines de l’analyse et la génération automatique de musique. Au-delà de leur performance, le présent projet vise à étudier le fonctionnement interne de deux de ces modèles, les plongements de mots et les transformeurs, ainsi que leur aptitude à s’adapter à des données musicales plutôt que textuelles. Ces expériences bénéficient à notre maîtrise de ces outils adaptés à la musique et contribuent à clarifier de nombreux parallèles entre le langage naturel et le langage musical.

National audience; Faced with the limits of representative democracy, digital public consultations provide an opportunity for citizens to contribute their opinions and ideas and for policy makers to involve the population more closely in the public decision making process. The design and deployment of such public consultations pose well-known issues related to potential biases in the questions or in the representativeness of the participants. In this article, we consider the novel issues that arise from the use of artificial intelligence methods to automatically analyze contributions in natural language. Conducting such analyses constitutes a difficult problem for which many approaches (relying on various assumptions and models) exist. Considering the responses to the open-ended questions of the French "Grand Débat National" as a case study, we show that it is impossible to reproduce the results of the official analysis commissioned by the government. In addition, we identify a number of implicit and arbitrary choices in the official analysis that cast doubts on some of its results. We show also that different methods can lead to different conclusions. Our study highlights the need for greater transparency in the automatic analyses of public consultations so as to ensure reproducibility and public confidence in their results. We conclude with suggestions for improving digital public consultations and their analysis so that they encourage participation and become useful tools for public debate.; Face aux limites de la démocratie représentative, les consultations numériques participatives publiques permettent de solliciter, à différents niveaux de pouvoir, des contributions de citoyens pour essayer de mieux impliquer les individus dans les décisions politiques. Leur conception et leur mise en place posent des problèmes bien connus tels que les biais dans les questions ou la représentativité de la population participante. Nous considérons, dans cet article, les problèmes nouveaux liés à l'utilisation de méthodes issues de l'intelligence artificielle pour l'analyse automatique de contributions en langage naturel. Réaliser une telle analyse est un problème difficile pour lequel il existe de nombreuses méthodes reposant sur des hypothèses et des modèles variés. En considérant comme cas d'étude les contributions aux questions ouvertes du Grand Débat National, nous montrons qu'il est impossible de reproduire les résultats de l'analyse officielle commandée par le gouvernement. En outre, nous identifions des choix arbitraires non explicités dans l'analyse officielle qui conduisent à émettre des doutes sur certains de ses résultats. Nous montrons également que différentes méthodes peuvent mener à des conclusions différentes. Notre étude met ainsi en lumière la nécessité d'une plus grande transparence dans les analyses automatiques de consultations ouvertes pour assurer leur reproductibilité et la confiance du public dans leur restitution. Nous concluons par des pistes d'amélioration des consultations participatives et de leur analyse pour qu'elles puissent inciter à la participation et être des outils utiles au débat public.

International audience

International audience; The efficiency of graph-based semi-supervised algorithms depends on the graph of instances on which they are applied. The instances are often in a vectorial form before a graph linking them is built. The construction of the graph relies on a metric over the vectorial space that help define the weight of the connection between entities. The classic choice for this metric is usually a distance measure or a similarity measure based on the euclidean norm. We claim that in some cases the euclidean norm on the initial vectorial space might not be the more appropriate to solve the task efficiently. We propose an algorithm that aims at learning the most appropriate vectorial representation for building a graph on which the task at hand is solved efficiently.

International audience; The link prediction problem for graphs is a binary classification task that estimates the presence or absence of a link between two nodes in the graph. Links absent from the training set, however, cannot be directly considered as the negative examples since they might be present links at test time. Finding a hard decision boundary for link prediction is thus unnatural. This paper formalizes the link prediction problem from the flexible perspective of preference learning: the goal is to learn a preference score between any two nodes---either observed in the network at training time or to appear only later in the test---by using the feature vectors of the nodes and the structure of the graph as side information. Our assumption is that the observed edges, and in general, shortest paths between nodes in the graph, can reinforce an existing similarity between the nodes feature vectors. We propose a model implemented by a simple neural network architecture and an objective function that can be optimized by stochastic gradient descent over appropriate triplets of nodes in the graph. Our first preliminary experiments in small undirected graphs show that our learning algorithm outperforms baselines in real networks and is able to learn the correct distance function in synthetic networks.

With the rapid expansion in the use of computers for producing digitalized textual documents, the need of automatic systems for organizing and retrieving the information contained in large databases has become essential. In general, information retrieval systems rely on a formal description or representation of documents enabling their automatic processing. In the most common representation, the so-called bag-of-words, documents are represented by the words composing them and two documents (or a user query and a document) are considered similar if they have a high number of co-occurring words. In this representation, documents with different, but semantically related terms will be considered as unrelated, and documents using the same terms but in different contexts will be seen as similar. It arises quite naturally that information retrieval systems can use the huge amount of existing textual documents in order to “learn”, as humans do, the different uses of words depending on the context. This information can be used to enrich documents’ representation. In this thesis dissertation we develop several original machine learning approaches which attempt at fulfilling this aim. As a first approach to document representation we propose a probabilistic model in which documents are assumed to be issued from a mixture of distributions over themes, modeled by a hidden variable conditioning a multinomial distribution over words. Simultaneously, words are assumed to be drawn from a mixture of distributions over topics, modeled by a second hidden variable dependent on the themes. As a second approach, we proposed a neural network which is trained to give a score for the appropriateness of a word in a given context. Finally we present, a multi-task learning approach, which is trained jointly to solve an information retrieval task, while learning on unlabeled data to improve its representation of documents, Avec l’essor de l’usage des ordinateurs pour la création de documents textuels digitalisés, le besoin de systèmes automatiques pour la recherche et l’organisation de l’information contenue dans de grandes bases de données est devenu central. En général, les systèmes de recherche d’information s’appuient sur une description formelle (ou représentation) des documents permettant leur traitement automatique.Dans la plus commune des représentations, appelée sac-de-mots, les documents sont représentés par l’ensemble des mots les constituant. Deux documents (ou bien un document et une requête) sont considérés comme similaires s’ils ont un grand nombre de mots en commun.Il est raisonnable de penser que les systèmes de recherche d’information devraient pouvoir utiliser les grandes quantités de données textuelles disponibles pour “apprendre”, à la façon des humains, les différents emplois d’un mot en fonction de son contexte. Cette information devrait pouvoir être utilisée pour enrichir la représentation des documents.Dans cette thèse, nous développons plusieurs approches originales d’apprentissage automatique qui tentent d’atteindre ce but.Comme première approche pour la représentation de documents nous proposons un modelé probabiliste qui suppose que les documents sont tirés d’un mélange de distributions sur des “thèmes”, représentés par une variable cachée qui conditionne une distribution multinomiale sur les mots. Simultanément, ce modèle suppose que les mots sont tirés d’une distribution sur les “sujets”, représentés quant à eux par une seconde variable cachée dépendante des thèmes.Comme deuxième approche, un réseau de neurones est proposé. Il est entraîné à donner un score reflétant le fait qu’un mot soit plus ou moins approprié étant donné un contexte.Finalement, nous présentons une approche multitâche entraînée de façon à résoudre conjointement une tâche de recherche d’information et tout en enrichissant la représentation des documents par l’utilisation des données non-étiquetées.

Statistical models for networks have been typically committed to strong prior assumptions concerning the form of the modeled distributions. Moreover, the vast majority of currently available models are explicitly designed for capturing some specific graph properties (such as power-law degree distributions), which makes them unsuitable for application to domains where the behavior of the target quantities is not known a priori. The key contribution of this paper is twofold. First, we introduce the Fiedler delta statistic, based on the Laplacian spectrum of graphs, which allows to dispense with any parametric assumption concerning the modeled network properties. Second, we use the defined statistic to develop the Fiedler random field model, which allows for efficient estimation of edge distributions over large-scale random networks. After analyzing the dependence structure involved in Fiedler random fields, we estimate them over several real-world networks, showing that they achieve a much higher modeling accuracy than other well-known statistical approaches.

International audience; Statistical models for networks have been typically committed to strong prior assumptions concerning the form of the modeled distributions. Moreover, the vast majority of currently available models are explicitly designed for capturing some specific graph properties (such as power-law degree distributions), which makes them unsuitable for application to domains where the behavior of the target quantities is not known a priori. The key contribution of this paper is twofold. First, we introduce the Fiedler delta statistic, based on the Laplacian spectrum of graphs, which allows to dispense with any parametric assumption concerning the modeled network properties. Second, we use the defined statistic to develop the Fiedler random field model, which allows for efficient estimation of edge distributions over large-scale random networks. After analyzing the dependence structure involved in Fiedler random fields, we estimate them over several real-world networks, showing that they achieve a much higher modeling accuracy than other well-known statistical approaches.

With the rapid expansion in the use of computers for producing digitalized textual documents, the need of automatic systems for organizing and retrieving the information contained in large databases has become essential. In general, information retrieval systems rely on a formal description or representation of documents enabling their automatic processing. In the most common representation, the so-called bag-of-words, documents are represented by the words composing them and two documents (or a user query and a document) are considered similar if they have a high number of co-occurring words. In this representation, documents with different, but semantically related terms will be considered as unrelated, and documents using the same terms but in different contexts will be seen as similar. It arises quite naturally that information retrieval systems can use the huge amount of existing textual documents in order to “learn”, as humans do, the different uses of words depending on the context. This information can be used to enrich documents’ representation. In this thesis dissertation we develop several original machine learning approaches which attempt at fulfilling this aim. As a first approach to document representation we propose a probabilistic model in which documents are assumed to be issued from a mixture of distributions over themes, modeled by a hidden variable conditioning a multinomial distribution over words. Simultaneously, words are assumed to be drawn from a mixture of distributions over topics, modeled by a second hidden variable dependent on the themes. As a second approach, we proposed a neural network which is trained to give a score for the appropriateness of a word in a given context. Finally we present, a multi-task learning approach, which is trained jointly to solve an information retrieval task, while learning on unlabeled data to improve its representation of documents, Avec l’essor de l’usage des ordinateurs pour la création de documents textuels digitalisés, le besoin de systèmes automatiques pour la recherche et l’organisation de l’information contenue dans de grandes bases de données est devenu central. En général, les systèmes de recherche d’information s’appuient sur une description formelle (ou représentation) des documents permettant leur traitement automatique.Dans la plus commune des représentations, appelée sac-de-mots, les documents sont représentés par l’ensemble des mots les constituant. Deux documents (ou bien un document et une requête) sont considérés comme similaires s’ils ont un grand nombre de mots en commun.Il est raisonnable de penser que les systèmes de recherche d’information devraient pouvoir utiliser les grandes quantités de données textuelles disponibles pour “apprendre”, à la façon des humains, les différents emplois d’un mot en fonction de son contexte. Cette information devrait pouvoir être utilisée pour enrichir la représentation des documents.Dans cette thèse, nous développons plusieurs approches originales d’apprentissage automatique qui tentent d’atteindre ce but.Comme première approche pour la représentation de documents nous proposons un modelé probabiliste qui suppose que les documents sont tirés d’un mélange de distributions sur des “thèmes”, représentés par une variable cachée qui conditionne une distribution multinomiale sur les mots. Simultanément, ce modèle suppose que les mots sont tirés d’une distribution sur les “sujets”, représentés quant à eux par une seconde variable cachée dépendante des thèmes.Comme deuxième approche, un réseau de neurones est proposé. Il est entraîné à donner un score reflétant le fait qu’un mot soit plus ou moins approprié étant donné un contexte.Finalement, nous présentons une approche multitâche entraînée de façon à résoudre conjointement une tâche de recherche d’information et tout en enrichissant la représentation des documents par l’utilisation des données non-étiquetées.

Documents are usually represented in the bag-of-word space. However, this representation does not take into account the possible relations between words. We propose here a graphical model for representing documents: the Theme Topic Mixture Model (TTMM). This model assumes two types of relations among textual data. Topics link words to each other and Themes gather documents with particular distribution over the topics. This paper defines the TTMM, compares it to the related Latent Dirichlet Allocation (LDA) model (Blei, 2003) and reports some interesting empirical results.

Although non-parametric tests have already been proposed for that purpose, statistical significance tests for non-standard measures (different from the classification error) are less often used in the literature. This paper is an attempt at empirically verifying how these tests compare with more classical tests, on various conditions. More precisely, using a very large dataset to estimate the whole ``population'', we analyzed the behavior of several statistical test, varying the class unbalance, the compared models, the performance measure, and the sample size. The main result is that providing big enough evaluation sets non-parametric tests are relatively reliable in all conditions.