ABSTRACT. The surge in social media use over the last decade brought a host of unintended and unanticipated complications, such as disinformation, polarization, and undisclosed content monetization. Some of these problems can be traced back to the lack of regulation governing the digital interactions, algorithmic decisions, monetization, and business strategies that shape the social media landscape. In an effort to address these issues, the European Union has recently approved the Digital Services Act (DSA), aiming to better regulate online spaces, including social media platforms. This talk will delve into challenges and opportunities that the implementation and enforcement of the DSA may bring for computational social scientists.

ABSTRACT. Community detection allows understanding how networks are organised. Ranging from social, technological, information or biological networks, many real-world networks exhibit a community structure. Consensual community detection fixes some of the issues of classical community detection like non-determinism. This is often done through what is called a consensus matrix. We show that this consensus matrix is not filled with relevant information only, it is noisy. We then show how to filter out some of the noise and how it can benefit existing algorithms.

ABSTRACT. Network community detection often relies on optimizing partition quality functions, like modularity or blockmodel likelihood. This optimization appears to be a complex unsupervised learning problem traditionally relying on various heuristic algorithms, which often fail to reach the optimal partition, and, therefore, may require further fine-tuning. We propose a new deep learning unsupervised model which consists of a two-layer bi-partite convolutional graph neural network, stacked with a fully connected attention vanilla neural network. The model can be used to fine-tune network partitions with respect to other quality/objective functions, such as block model likelihood or description length. Furthermore, it can be used for supervised community detection, where one seeks to learn how to extrapolate the community structure provided for a certain part of the network to the rest of the network nodes. Finally, the configuration of the model depending only on the selected dimensionality of the node embedding but not on the dimensionality of the network nor the number of network communities, makes it suitable for transferring pre-trained model architectures and parameters between different networks. It also enables applying ensembles of the pre-trained transferred models to improve performance over the target network.

ABSTRACT. Numerous real-world question correspond to the vertex cover problem, which consists in finding a subset of nodes that touch all the edges of a graph. As this problem is hard to solve exactly, approximate solutions are used when handling large complex networks. Although their results may be excellent, there is mathematical guarantee of how far it is from optimum. In this work, we propose a method that certifies the quality of a vertex cover of a given network. The certificate is given by an approximate result and a lower-bound on the minimum value. Tested on 114 real-world networks with up to three billion edges, our method certifies that the results of state-of-the-art heuristics for vertex cover is within 1% of the optimum value on two thirds of the networks. This work shows that valuable quality certificates can be given for existing heuristics on specific instances without loosing on scalability. As it may generalise to other algorithmic problems, it opens a door for further research and for deployment in real-world applications.

ABSTRACT. Networks have provided a representation for a wide range of real systems. Uncovering densely connected groups in these networks is the goal of community detection. Communities represent fundamental structures for understanding the organization of real-world networks. This paper lays the foundation for an application of community detection on transactional network focused on finding the most effective way of simplifying multigraphs to weighted graphs. We tested different weights' calculation function and community detection algorithms. A comparison of the outputs based on extrinsic and intrinsic evaluation metrics is then held.

ABSTRACT. This paper positions itself in the stream of literature that complements the classical theories of knowledge transfer, acquisition, and absorption with the studies of networks and their properties. The novelty that this paper brings is twofold. On the one hand, by relying on the concept of Simmelian ties (Simmel, 1950), we further the discussion regarding the role that the strength of a tie plays in the transfer of knowledge and information across a network. In particular, by considering the organizational structure of a company and the personal relationships of the employees of the company, we want to study the transfer of information between subjects belonging to two different hierarchical levels and the role that Simmelian ties play. On the other hand, we enrich the classical innovation diffusion (Bass, 1969) and opinion dynamics (DeGroot, 1974) models with a characterization that is linked with the structural characteristics of the network of the population analyzed. Eventually, we propose a model that accounts for the structural characteristics under discussion.

ABSTRACT. This work intends to address the process of diffusion of AI in neuroscience. The corresponding underlying dynamics can be captured through both the evolution of the disciplinary ecosystem around the whole neuroscience and the structure of collaborations among the scientists involved in this field of research. To do this, we build a bibliometric corpus including all neuroscience articles published between 1940 and 2019, that we extracted from the Microsoft Academic Knowledge Graph database. Then we explore specifically the development of a dedicated AI specialty in neuroscience with its own scientific community by conducting analyses of both the egocentric citation network around these neuroscience articles and the associated co-authorship network. We show especially with the first network a progressive specialization of references toward computer science, mathematics and engineering, while its impact is more broadly distributed across the entirety of the neuroscience field (see Fig.1A), mainly into engineering, radiology and neuroimaging technologies for clinical research and medicine. With the time-aggregated co-authorship network, including the main collaborations since 1970, we show that a small set of researchers (around 1.3% of them) who together have authored the most publications in AI research in peer-reviewed journals oriented toward computational neuroscience, mostly based on neural networks, and that does not maintain links with the rest of neuroscience community which does not publish AI research. These AI specialists exhibit also particular disciplinary backgrounds and academic trajectories that are contrasting with the typical ones encountered in neuroscience. These results have led us to qualify the AI scientists as ‘outsiders’ of neuroscience. According to the state of the art of the formation of a scientific specialty encountered in science studies, this work thus shows through AI in neuroscience a pattern of diffusion of knowledge in a scientific field of research, namely a formation and a transformation of a special AI-research alongside neuroscience, with its own evolving scientific community and bibliographic references, and which seems though to contribute to the main challenges of this field of research. This preliminary work is thus a part of a larger study of the history of neuroscience through AI.

ABSTRACT. As we showed in [1], the typical property of online recommendation algorithms of reinforcing the attention on the latest trends can be responsible for the ever faster rise and fall of collective attention around online objects (i.e. YouTube videos), an effect that we call “junk news bubbles”. In an accelerating society, it is not only the consumption of online information that is becoming more ephemeral: a similar phenomenon has been observed for the citation cycles of scientific production [2]. In the case of scientific production, however this acceleration may depend less on specific recommendation algorithms than on the practices of researchers for linking their work to the existing literature, the “shoulders of giants” on which the scientific innovation is supposed to stand. In this context, indeed, other phenomena concerning the temporality of research has been observed: the consolidations of canons and the slowing down of disruptiveness [3,4]. We start to analyze these two apparently opposite phenomena, the acceleration of attention cycles and crystallization of canons, applying a varied set of statistical measures to several corpora, at different scales of granularity (from precise thematic domains, like i.e., astrophysics or cognitive science or artificial intelligence, to large domains identified in the Web of Science, i.e., life science or technology). We observe the two effects in all our corpora, suggesting that this double effect can be indeed associated with a super-cumulative advantage process promoting again and again the same classic papers and penalizing most of the novelties, that become more ephemeral. We build therefore a simple model to reproduce the citation patterns and we show that this effect, without any other exogeneous recommendation forcing, can be reconducted to the exponential growth of the scientific literature. We also analyze the impact of this effect on innovation cycles.

[1] Castaldo, M., Venturini, T., Frasca, P., & Gargiulo, F. (2020). Junk news bubbles modelling the rise and fall of attention in online arenas. 1461444820978640. [2] Lorenz-Spreen, Philipp, et al. "Accelerating dynamics of collective attention." 10.1 (2019): 1-9. [3] Chu, Johan SG, and James A. Evans. "Slowed canonical progress in large fields of science." Proceedings of the National Academy of Sciences [4] Park, Michael, Erin Leahey, and Russell J. Funk. "Papers and patents are becoming 118.41 (2021). less disruptive over time." Nature613.7942 (2023): 138-144.

ABSTRACT. Rankings are a well-established tool to evaluate the performance of actors in different sectors of the economy, and their use is increasing even in the context of the startup ecosystem, both on a regional and on a global scale. Although rankings meet the demand for measurability and comparability, they often provide an oversimplified picture of the status quo, which, in particular, overlooks the variability of the socio-economic conditions in which the quantified results are achieved. In this paper, we describe an approach based on constructing a network of world countries, in which links are determined by mutual similarity in terms of development indicators. Through the instrument of community detection, we perform an unsupervised partition of the considered set of countries, aimed at interpreting their performance in the StartupBlink rankings. We consider both the global ranking and the specific ones (quality, quantity, business). After verifying if community membership is predictive of the success of a country in the considered ranking, we rate country performances in terms of the expectation based on community peers. We are thus able to identify cases in which performance is better than expected, providing a benchmark for countries in similar conditions, and cases in which performance is below the expectation, highlighting the need to strengthen the innovation ecosystem.

ABSTRACT. Measures of citations, fame, and popularity are used frequently as proxies for the quality of scientific and cultural work to study the dynamics of success, attention, and collective memory. The advent of the internet has made it possible to estimate popularity online by measuring the overall quantity of attention through searches, video reproductions, and page views. Yet, raw measures of attention often conflate multiple forms of popularity. To address this issue, we unpack popularity in a framework focused on multiple dimensions: (i) volume, (ii) longevity, (iii) connectivity, and (iv) globality. These dimensions are computed from the page views and demographic information on Wikipedia of the 100k most famous people collected by the Pantheon database in 2020. We find that globality, connectivity, and longevity explain 66%, 20%, and 3% of the variance in volume, respectively. Taken together, the globality, connectivity, and longevity dimensions explain 76% of volume, after controlling for individual characteristics such as occupation, gender, age and nationality. Our findings take one step further in mapping the historical geography of fame and provide a more nuanced and comprehensive framework for measuring popularity.

ABSTRACT. The ocean plays a central role in the Earth system. It regulates geophysical processes and is thus crucial to understand the pace of global change. It is also the largest ecosystem and thus host of biodiversity. Understanding the interactions among physical, natural, and human processes, how they evolve in time and how they participate in the functioning of ecosystems is thus of crucial importance. The development of novel techniques under the umbrella of Big Data analytics and Data science offers new opportunities for complex systems to analyze the Ocean system. We will present opportunities and challenges for complex systems in this context and in particular recent advances in the analysis of the largest multi-taxa marine megafauna tracking dataset recently assembled.

ABSTRACT. A container terminal is a complex system where different service level conditions are required for different vessels and customers. Thus, simulating terminal processes is often a complex task requiring a modeling approach that facilitates decision analysis. In this study, we first propose to model the berth allocation (BAP) integrated with the quay crane assignment (QCAP). A new mixed integer bi-objective linear program is proposed to reduce the dwell time of each ship, the penalty by late ships, and to find the optimal number of quay cranes (QCs) needed per ship. Then, we develop a resolution procedure based on the non-dominated genetic sorting algorithm (NSGA-III). We also use statistical analysis to identify the control parameters of (NSGA-III). Then, we implement the algorithm calibrated with the obtained control parameters. We conduct a computational experiment on a set of large randomly generated instances to highlight the benefits and suitability of the proposed approach. The numerical results show the efficiency of the approach.

ABSTRACT. A fundamental assumption in economics is that rational individuals act in their own self interest. One implication is that, when trading, buyers are supposed to seek for the lowest price and sellers for the highest one and social interactions are not considered. It is now largely accepted that social relationships affect the efficiency of a market structure (centralized or decentralized) (Babus et al. 2013, Opp & Glode 2016, Glode & Opp 2017). The objectives of the current study is to examine the network structures of a very specific market : the Boulogne-sur-mer fish market. On this market two market struc- tures coexist, each beeing used by the same buyers and sellers, exchanging similar goods. The two submarkets are a centralized one (Auctions) and a decentralized one (over-the-counter market). For each sub-market we examine (1) the global network structure, (2) the local network structure, and (3) we identify the traders characteristics that best explain the network structures. by comparing the results, we can compare the role of trust (bilateral market) and reputation (auction market) in the individual choices of trading partners. Structural measures are used to characterize networks structures. Exponential ran- dom graph models are used to evaluate how trader characteristics explain purchasing patterns, and how the influence of these characteristics vary with the market mecha- nism. We bring into the light that, when the transaction links on the auction market re- flects the economic constraints of the partners, the relationships on the bilateral market depends on something more. Clearly, the prices of the bilateral transactions are the con- sequences of economics and non economics determinants. At first glance, the stable co-existence of two market structures looks like a paradox. Our results help to under- stand the distinctive characteristics and functioning of each sub-market. This discussion contributes to the debate about the efficiency of market structures.

ABSTRACT. Fans, sports organizations, as well as players use social networks like Twitter to build and maintain their identity and sense of community. Breaking news about soccer sometimes comes first on Twitter than traditional media channels and provides an excellent means to access instantaneous information from official and unofficial sources. Soccer has more than 3.5 billion fans worldwide, and it is estimated that 1.3 billion of them are females (around 38\%). Our work examines whether, in a male-dominated environment, women and men differ in communication patterns. Women soccer fans tend to experience biases and prejudices, and our question relies on how this is translated to online spaces. Through our study, we look into the patterns of interaction between men and women and how communication evolves over 3 months (March 7 to June 6, 2022) for English and Portuguese tweets. After our data preprocessing, we have 7,676,624 tweets in English (6,365,239 are from males and 1,313,731 are from females) and 2,958,443 tweets in Portuguese (2,312,415 are from males and 648,395 from females). We find that the Portuguese network has a higher women ratio in interactions and a lower homophily than the English network. The network structure from women's tweets tends to be denser, with higher average clustering and lower assortativity than men's one. We also observe that, in soccer, women tend to express higher levels of joy and anticipation than men in both languages, and disgust tends to be more gender-neutral with slightly higher levels for males. Interestingly, we did not find any qualitative difference in relation to the gender differences in emotion between the English and Portuguese collected tweets. We thus found that the emotional response across genders seems independent of the overall network structure.