Benjamin Renoust

Cette thèse présente une nouvelle méthodologie pour analyser des réseaux. Nous développons l'intrication d'un réseau multiplex, qui se matérialise sous forme d'une mesure d'intensité et d'homogénéité, et d'une abstraction, le réseau d'interaction des catalyseurs, auxquels sont associés des indices d'intrication. Nous présentons ensuite la mise en place d'outils spécifiques pour l'analyse visuelle des réseaux complexes qui tirent profit de cette méthodologie. Ces outils présente une vue double de deux réseaux,qui inclue une un algorithme de dessin, une interaction associant brossage d'une sélection et de multiples liens pré-attentifs. Nous terminons ce document par la présentation détaillée d'applications dans de multiples domaines.

Analysis of cultural heritages, particularly their construction years, provides new insights into human history. However, due to natural disasters, wars, material deterioration, and human errors, records documenting the construction years of many artifacts have often been lost. Historians and experts can estimate construction years within specific ranges using chemical-based analysis technologies or extensive historical research. Given the vast number of collected artifacts, applying these conventional methods to every artifact is impractical. To address this challenge, we developed a deep neural network model designed for Buddha statues to estimate an artifact’s construction year from its image. One major challenge in this task is the heterogeneity of the labels: the training samples include both precise construction years and possible ranges (e.g., a dynasty or a century) estimated by historians. To unify these heterogeneous labels during training, we represent them as probabilistic distributions. In our previous work Qian et al. (2021), we assumed that the ambiguity in heterogeneous construction year labels followed a Gaussian distribution, assigning the highest likelihood to the midpoint of the designated time range. However, this assumption does not always hold. In this paper, we propose representing heterogeneous construction year labels as a uniform distribution, assigning equal probability to all points within the designated time range. Based on this label representation, we designed a semi-supervised learning loss function to leverage both labeled and unlabeled samples during training. Our experimental results demonstrate that our method achieves a mean absolute error of 34.3 years on a test set consisting of Buddha statues constructed between 400 and 1403. These results are further analyzed in two ways. First, we compared our model’s performance to the image quality BRISQUE score, revealing a correlation between higher image quality and lower prediction error rates. Second, we validated our predictions with experts, assessing the level of agreement with our model, the challenges in determining construction years, and identifying features of interest in the artifacts.

Graph distance measures have emerged as an effective tool for evaluating the similarity or dissimilarity between graphs. Recently, there has been a growing trend in the application of movie networks to analyze and understand movie stories. Previous studies focused on computing the distance between individual characters in narratives and identifying the most important ones. Unlike previous techniques, which often relied on representing movie stories through single-layer networks based on characters or keywords, a new multilayer network model was developed to allow a more comprehensive representation of movie stories, including character, keyword, and location aspects. To assess the similarities among movie stories, we propose a methodology that utilizes a multilayer network model and layer-to-layer distance measures. We aim to quantify the similarity between movie networks by verifying two aspects: (i) regarding many components of the movie story and (ii) quantifying the distance between their corresponding movie networks. We tend to explore how five graph distance measures reveal the similarity between movie stories in two aspects: (i) finding the order of similarity among movies within the same genre, and (ii) classifying movie stories based on genre. We select movies from various genres: sci-fi, horror, romance, and comedy. We extract movie stories from movie scripts regarding character, keyword, and location entities to perform this. Then, we compute the distance between movie networks using different methods, such as the network portrait divergence, the network Laplacian spectra descriptor (NetLSD), the network embedding as matrix factorization (NetMF), the Laplacian spectra, and D-measure. The study shows the effectiveness of different methods for identifying similarities among various genres and classifying movies across different genres. The results suggest that the efficiency of an approach on a specific network type depends on its capacity to capture the inherent network structure of that type. We propose incorporating the approach into movie recommendation systems.

In automatic art analysis, models that besides the visual elements of an artwork represent the relationships between the different artistic attributes could be very informative. Those kinds of relationships, however, usually appear in a very subtle way, being extremely difficult to detect with standard convolutional neural networks. In this work, we propose to capture contextual artistic information from fine-art paintings with a specific ContextNet network. As context can be obtained from multiple sources, we explore two modalities of ContextNets: one based on multitask learning and another one based on knowledge graphs. Once the contextual information is obtained, we use it to enhance visual representations computed with a neural network. In this way, we are able to (1) capture information about the content and the style with the visual representations and (2) encode relationships between different artistic attributes with the ContextNet. We evaluate our models on both painting classification and retrieval, and by visualising the resulting embeddings on a knowledge graph, we can confirm that our models represent specific stylistic aspects present in the data.

While Buddhism has spread along the Silk Roads, many pieces of art have been displaced. Only a few experts may identify these works, subjectively to their experience. The construction of Buddha statues was taught through the definition of canon rules, but the applications of those rules greatly varies across time and space. Automatic art analysis aims at supporting these challenges. We propose to automatically recover the proportions induced by the construction guidelines, in order to use them and compare between different deep learning features for several classification tasks, in a medium size but rich dataset of Buddha statues, collected with experts of Buddhism art history.

Real world complex networks often exhibit multiplex structure, connecting entities from different aspects of physical systems such as social, transportation and biological networks. Little is known about general properties of such networks across disciplines. In this work, we first investigate how consistent are connectivity patterns across 35 real world multiplex networks. We demonstrate that entanglement homogeneity and intensity, two measures of layer consistency, indicate apparent differences between social and biological networks. We also investigate trade, co-authorship and transport networks. We show that real networks can be separated in the joint space of homogeneity and intensity, demonstrating the usefulness of the two measures for categorization of real multiplex networks. Finally, we design a multiplex network generator, where similar patterns (as observed in real networks), are emerging over the analysis of 11{,}905 synthetic multiplex networks with various topological properties.

In computer vision, visual arts are often studied from a purely aesthetics perspective, mostly by analysing the visual appearance of an artistic reproduction to infer its style, its author, or its representative features. In this work, however, we explore art from both a visual and a language perspective. Our aim is to bridge the gap between the visual appearance of an artwork and its underlying meaning, by jointly analysing its aesthetics and its semantics. We introduce the use of multi-modal techniques in the field of automatic art analysis by 1) collecting a multi-modal dataset with fine-art paintings and comments, and 2) exploring robust visual and textual representations in artistic images.

Network models have been increasingly used in the past years to support summarization and analysis of narratives, such as famous TV series, books and news. Inspired by social network analysis, most of these models focus on the characters at play. The network model well captures all characters interactions, giving a broad picture of the narration's content. A few works went beyond by introducing additional semantic elements, always captured in a single layer network. In contrast, we introduce in this work a multilayer network model to capture more elements of the narration of a movie from its script: people, locations, and other semantic elements. This model enables new measures and insights on movies. We demonstrate this model on two very popular movies.

Online communities are used across several fields of human activities, as environments for large-scale collaboration. Most successful ones employ professionals, sometimes called “community managers” or “moderators”, for tasks including onboarding new participants, mediating conflict, and policing unwanted behaviour. Network scientists routinely model interaction across participants in online communities as social networks. We interpret the activity of community managers as (social) network design: they take action oriented at shaping the network of interactions in a way conducive to their community’s goals. It follows that, if such action is successful, we should be able to detect its signature in the network itself. Growing networks where links are allocated by a preferential attachment mechanism are known to converge to networks displaying a power law degree distribution. Growth and preferential attachment are both reasonable first-approximation assumptions to describe interaction networks in online communities. Our main hypothesis is that managed online communities are characterised by in-degree distributions that deviate from the power law form; such deviation constitutes the signature of successful community management. Our secondary hypothesis is that said deviation happens in a predictable way, once community management practices are accounted for. If true, these hypotheses would give us a simple test for the effectiveness of community management practices. We investigate the issue using (1) empirical data on three small online communities and (2) a computer model that simulates a widely used community management activity called onboarding. We find that onboarding produces in-degree distributions that systematically deviate from power law behaviour for low-values of the in-degree; we then explore the implications and possible applications of the finding.