Wednesday Lightnings

Complex systems often experience perturbations and disturbances, from overload failures in power systems to species extinction in ecological networks. The impact of such perturbations is often subtle, the system exhibits a minor response, but continues to sustain its global functionality. However, in extreme cases, a large enough perturbation may lead to a large-scale collapse, with the system abruptly transitioning from a functional to a dysfunctional dynamic state. For instance, in cellular dynamics, genetic knockouts, beyond a certain threshold, lead to cell death; in ecological systems, changes in environmental conditions may, in extreme cases, cause mass-extinction; and in infrastructure networks, a cascading failure, at times, results in a major blackout. When such collapse occurs, the naïve instinct is to reverse the damage, retrieve the failed nodes and reconstruct the lost links. Such response however is seldom efficient, as (i) we rarely have access to all system components; (ii) even if we could reverse the damage, due to hysteresis, in many cases, the system will not spontaneously regain its lost functionality. To address this challenge, we consider here a two-step recovery process: Step I. Restructuring. Retrieving the network topology to a point where the system can potentially regain functionality • Step II. Reigniting. Introducing dynamic interventions to steer the system back to its functional state. The challenge in Step II is that we are typically limited in our potential interventions. Hence, we seek to reignite the system via micro-interventions, e.g., controlling just a single or, at most, few nodes. Therefore, in this talk, we present single-node reigniting, unveiling a new dynamic phase of complex systems, in which a failed system can be revived by controlling just a microscopic set of nodes. This helps jolt the network back to functionality – offering a practical network dynamics defibrillator.

See attachment If accepted, I (FB) will be the presenter. This is the only contribution I am submitting as presenter. My other contributions, if accepted, will be presented by co-authors.

Since real-world epidemics are very likely characterized by non-exponential infections and curings, non-Markovian models are expected to describe real epidemic processes better. In order to account for memory in Markovian stochastic processes, we replace the standard differential operator in the Markovian equations with the well-known Caputo fractional derivative, an operator which generalizes the notion of standard derivative to non-integer orders (i.e., 0 < α < 1) and which incorporates memory effects “summing” over the past. We apply it to generalize the equations describing Markovian epidemic processes on networks and we prove that is possible to define a microscopic epidemic description that is compatible with the fractional framework. Without making any assumption on the memory structure of the fractional process we obtain bounds for the average fraction of infected individuals (i.e., prevalence) and we show that for 0 < α < 1 the fractional epidemics is always worse (i.e. more dangerous) than the Markovian epidemics because grows faster in the beginning and survives more in the long-run. Assuming that the fractional process satisfies the semi-Markov property, which means that the memory of the process is reset every time there is an update in the viral state of the network, we analytically prove that is possible to define a novel microscopic epidemic description that is compatible with the fractional Chapman-Kolmogorov equation. Finally we also show under which conditions any fractional stochastic process can be reduced to a corresponding Markov process via different kinds of time transformations, providing a practical tool to extend the well-known theory of Markovian epidemic process to a more general framework.

Green industrial strategy comprises policies that help steer a country's economic structure towards sectors that are cleaner and produce green technologies. In both policy practice and the economic literature, green industrial strategy has been growing rapidly in popularity. In this study, we use network science methods to expose underlying patterns of industrial strategy that drive policy innovations. While network science is relatively novel to policy making, existing literature has stressed that policy making is a complex, evolving, path-dependent process. This paper is the first to study industrial strategy using network science and adds to a nascent literature on networks in policymaking. We introduce a new dataset of green industrial strategy, spanning policies between 2008--2022 that target the value chain of three major renewable energy technologies: solar pv, wind energy, and batteries for electric vehicles (EVs). Using this data, we create three separate co-occurrence networks of countries, products, and policy instruments. We find that industrial strategy is a path dependent process, and that the network structure is informative of policy innovation: countries have a tendency to introduce new policies if they are more aligned with the policy instruments that are already in place, the products that it already targets, or if related countries already have that policy in place. We show that the direction of green industrial policy can thus - to some extent - be predicted using our networks. One of the three networks is of industrial policy instruments, where instruments are connected if they are often co-applied to the same product in the same country. Fig. 1 overlays the network with a specific policy innovation in India, which was more than four times more likely to be introduced in 2017 than in a situation where no related policy instrument had already been introduced by India on that product before.

referring to the pdf file

Medulloblastoma (MB) is a rare childhood brain tumor historically stratified into four molecular subgroups: WNT, SHH, Group 3 (G3), and Group 4 (G4). While SHH and WNT are widely accepted as separate subgroups, G3 and G4 are currently viewed as a continuum within the non-SHH/non-WNT subgroup. The present work focuses on this subgroup continuum, as recent research on various omics datasets has suggested the existence of a putative, additional G3-G4 subgroup with intermediate characteristics. We present findings that identify and characterize the additional G3-G4 subgroup using two distinct omics datasets and two distinct methodologies. First, we characterized the putative subgroup based on genes' persistent association in multilayer network communities across multi-omics data, and identified patients that do not belong to either the G3 nor the G4 subgroups. Given data scarcity is a common problem in rare disease analyses, including in transcriptomic data, in our second procedure we sought to overcome this limitation by generating new, synthetic patients within the G3-G4 subgroup from a microarray study. By leveraging the Variational Autoencoder generative ability, we were able to overcome the problem of data scarcity and found quantitative support for the G3-G4 subgroup definition. Overall, we found a subset of genes that separate the G3-G4 subgroup, both through our multilayer network and synthetic-data approaches. These approaches show the potential of overcoming dimensionality limitations across different data types, enhancing rare disease subtyping tasks and allowing for a deeper understanding of the underlying biology.

Schumpeterian creative destruction most naturally relates innovation and its complement exnovation in a linear progression of realized possibilities from the past to the present. The phylogenetic tree is a structural variation that represents contingent paths of biological innovations to a species. For a technological capability, paths are manifold, or a “truss.” We propose a simple model for a population of agents innovating while outrunning exnovation to contrast the impact of dynamical parameters with structural ones. We calculate its phase diagram to show how dynamics and structural connectivity conspire to unleash innovative diversity or to drive it extinct.

The dominance of online social media data as a source for large-scale social network studies has recently been challenged by networks constructed from state-curated register data. In this paper, we investigate how the two compare, focusing the Dutch online social network (OSN) Hyves and a register-based social network (RSN) of the Netherlands. First and foremost, we find that the connectivity of the two population-scale networks is strikingly similar, especially between nearby municipalities, with more long-distance ties captured by the OSN. This result holds when correcting for population size and geographical distance, notwithstanding that these two factors appear to be the main drivers of connectivity. Second, we show that the community structure of neither network follows strict administrative geographical delineations (e.g., provinces). Instead, communities appear to either center around large metropolitan areas or, outside of the country's most urbanized area, are comprised of large blocks of interdependent municipalities. Beyond population and distance-related patterns, communities also highlight the persistence of deeply rooted sociocultural communities that can be linked to religion. The findings presented in this work aid in comparing results from future studies in which online social networks and register-based social networks are used to obtain insights into the social network structure of an entire population.

Random geometric graphs (RGGs) are often used to model real-life social, technological and biological networks. An artificial clique may signal anomaly or malicious interventions. We study the problem of detecting such a artificial (or, hidden, planted) clique in RGGs. Formally, we consider a RGG where n vertices are located at random on a finite two-dimensional torus; two vertices are connected if the distance between them is at most r. Then, k vertices are chosen at random to form a planted clique by adding the missing edges between them. Our goal is to design algorithms that find the planted clique with high probability (w.h.p). We analyze two algorithms. The vertex degree (VD) algorithm declares that the planted clique is the k vertices with highest degrees, just as in the earlier literature for the Erdös-Rényi (ER) random graphs. We prove that the VD algorithm works for k greater than log(n) times the average degree, similarly to the ER model. Interestingly, the largest clique in a RGG is much greater than that. Our main contribution is the common neighbor (CN) algorithm that works as follows. If a pair of vertices has exactly k-2 common neighbors, then CN declares that this pair and its common neighbors are the planted clique. We prove that the CN algorithm finds a planted clique w.h.p. in a large range of regimes, for instance (surprisingly!) when k is finite, while the average degree is of the order log(n). We conclude that even without knowing the positions of the vertices, we can leverage the hidden geometry to greatly enhance detection, in comparison to non-geometric models.

This work presents an experimental evaluation of algorithmic and neural community detection methods for temporal graphs, comparing more established approaches with newly introduced deep learning models designed for node-level clustering. We devise a novel stochastic block model-based approach 3] to generate time-varying attributed temporal graphs with community ground truths, and employ it to compare the performance of distinct state-of-the-art solutions. In this framework, node features are kept fixed, while community structures are dynamic — allowing to study how well differing solutions are able to jointly exploit the feature space and the temporal dynamics of the generated graphs to improve on the detectability threshold of communities and allow better recovering the hidden community structure of the generated networks. In addition, we assess the performance of the models on a set of diverse real-world datasets of various scales, focused on a transductive learning setting, and extending to an inductive learning setting when possible. The main motivation behind this work are the recent advances in network representation learning, which have also sparked a renewed interest in the improvement and development of new strategies for learning on spatio-temporal signals.

Complex networks are powerful tools for modeling real-world phenomena, including social interactions, economic transactions, and biological processes. Many of these phenomena are characterized by a continuous flow of interactions occurring at fast time scales, often modeled as stream graphs or link streams, frameworks specifically designed to handle fine-grained temporal data without unnecessary aggregation. Working within this framework necessitates extending methods developed for static networks, including the classic problem of community detection. In previous work we focused on the well-known modularity quality function for community structures in static networks and introduced an extension for link streams: the Longitudinal Modularity. This extension does not require a predefined time scale of analysis or any preprocessing of the natural temporal interactions in the network. In this work, we present an algorithm designed to identify time communities in link streams that maximize Longitudinal Modularity.