Abstract: Location-based services have become a popular subject of research over the past decade thanks to their significance as a novel source of geo-referenced data that provide solutions for a variety of research problems in a host of disciplines. Despite their richness in terms of the multiple data layers that become available, these datasets are often sparse and are characterised by skewed distributions which make the application of classical statistical frameworks and machine learning algorithms in this context a challenge. In this talk, we will review a wider range of data aggregation and coarse graining techniques that enable useful characterisations of complex location-based systems and find applicability in many real world applications.

Abstract: Many complex systems are modular, in that their components are organized in tightly-integrated subsystems that are weakly-coupled to one another. Modularity has been argued to be important for evolvability, state-space exploration, subsystem specialization, and many other important functions. The problem of decomposing a system into weakly-coupled modules, which has been studied extensively in graphs, is here considered in the domain of multivariate dynamics, a commonly-used framework for modeling complex physical, biological and social systems. We propose to decompose dynamical systems using the idea that modules constrain the spread of localized perturbations. We find partitions of system variables that maximize a novel measure called `perturbation modularity', defined as the auto-covariance of a coarse-grained description of perturbed trajectories. Our approach effectively separates the fast intra-modular from the slow inter-modular dynamics of perturbation spreading (in this respect, it is a generalization of the Markov stability method of community detection). Perturbation modularity can capture variation of modular organization across different system states, time scales, and in response to different kinds of perturbations. We argue that our approach offers a principled alternative to detecting graph communities in networks of statistical dependency between system variables (e.g. `relevance networks', 'functional networks', and other networks based on correlation or information-transfer measures). Using coupled logistic maps, we demonstrate that the method uncovers hierarchical modular organization encoded in a system's coupling matrix. Additionally, we use it to identify the onset of self-organized modularity in certain parameter regimes of homogeneously-coupled map lattices (originally popularized by Kaneko). Our approach offers a powerful and novel tool for exploring the modular organization of complex dynamical systems.

Abstract: In recent years, the increasing availability of publication records has attracted the attention of the scientific community. In particular, many efforts have been devoted to the study of the organization and evolution of science by exploiting the textual information extracted from the title and abstract of the articles. However, lesser attention has been devoted to the core of the article, i.e., its body. The access to the entire text, instead, paves the way to a better comprehension of the relations of similarity between articles. In the present work, concepts are extracted from the body of the scientific articles available on the ScienceWISE platform, and are used to build a network of similarity between articles. The resulting weighted network possesses a considerably high edge density, spoiling any attempt of associating communities of papers to specific topics. This happens because not all the concepts inside an article are truly informative and, even worse, they may not be useful to discriminate articles with different contents. Moreover, the presence of ``generic concepts'' with a loose meaning implies that a considerable amount of connections is made up by spurious similarities. To eliminate generic concepts, we introduce a method to evaluate the concepts' relevance according to an information-theoretic approach. The significance of a concept $c$ is defined in terms of the distance between its maximum entropy distribution, $S_{max}$, and the empirical one, $S_c$, calculated using the frequency of occurrence inside papers. By evaluation such distance, generic concepts are automatically identified as those with an entropy closer to the maximum. The progressive removal of generic concepts retaining only the ``meaningful'' ones has a twofold effect: it decreases sensibly the density of the network and reinforce meaningful relations. By applying different ``filtering thresholds,'' we unveil a refined topical organization of science in a coarse-grained way.

Abstract: In this talk I will discuss the problem of developing predictive ?models for layered data. Time-resolved networks are a typical example of layered data, since each time window results in a specific pattern of connections. I will present stochastic tensorial block models as a valid approach to predict missing information in network data with different layers of information. I will discuss results for two cases: a temporally resolved e-mail communication network and a drug-drug interaction network in different cell lines.

Abstract: We study a large evolving network obtained from Twitter created by a sample of users @-mentioning each other. We find that people who have potentially the largest communication reach (according to a dynamic centrality measure) use sentiment differently than the average user: for example they use positive sentiment more often and negative sentiment less often. Furthermore, we use several algorithms for community detection based on structure of the network and users' sentiment levels to identify several communities. These communities are structurally stable over a period of months. Their sentiment levels are also stable, and sudden changes in daily community sentiment in most cases can be traced to external events affecting the community. Based on our findings, we create and calibrate a simple agent-based model that is capable of reproducing measures of emotive response comparable to those obtained from the observed data.

Abstract: We present a framework for the construction and analysis of directed metabolic reaction graphs that can be tailored to reflect different environmental conditions. In the absence of information about the environmental context, we propose a Probabilistic Flux Reaction Graph (PRG) in which the weight of a connection between two reactions is the probability that a randomly chosen metabolite is produced by the source and consumed by the target. Using context-dependent flux distributions from Flux Balance Analysis (FBA), we produce a Flux-Balance Graph (FBG) with weighted links representing the amount of metabolite flowing from a source reaction to a target reaction per unit time. The PRG and FBG graphs are analyzed with tools from network theory to reveal salient features of metabolite flows in each biological context. We illustrate our approach with the directed network of the central carbon metabolism of Escherichia coli, and study its properties in four relevant biological scenarios. Our results show that both flow and network structure depend drastically on the environment: graphs produced from the same metabolic model in different contexts have different edges, components, and flow communities, capturing the biological re-routing of metabolic flows inside the cell. By integrating probabilistic and FBA-based analysis with tools from network science, our results provide a framework to interrogate cellular metabolism beyond standard pathway descriptions that are blind to the environmental context.