Skip to main content

Decomposition and Discovery of Complex Networks

Final Report Summary - DEDINET (Decomposition and Discovery of Complex Networks)

DEDINET has aimed at addressing the problem of unraveling the different mechanisms that are responsible for the complex interaction topology we observe in complex systems such as cells, societies or economies.

This project lies on the premise that for many real-world complex systems there is not a single mechanism of interaction between the components of a complex system. For instance, social network users establish interactions with a number of people because they are connected in one or more different levels (emotional, professional, cultural, etc).

Assuming that stochastic block models are good at capturing connectivity patterns, the overarching goal of DEDINET is to develop a framework that enables the identification of a set of orthogonal block models, and the discovery of complex networks from sparse/incomplete empirical data sets; and to apply this framework to problems in systems biology.

To achieve our goal, we have developed a comprehensive framework that is able to sample generative block models in networks that involve different types of nodes (such as bipartite graphs) and different types interactions (weighted, unweighted and categorical) (for instance, enzymes and reaction in metabolic networks or users and movies in a movie rating system).

As part of our research, we have applied the inference framework to different types of biological and social networks and shown that the block models we sample with our method are able to capture different uncorrelated properties of the nodes in the networks. This observation demonstrates that the hypothesis that there are several (overlapping) block models that generate the structure of the network is true at least in some cases.

Finally, we have developed an inference framework to extract sets of two block models that better describe the observed topology of a network that results from the aggregation of interactions at different layers. In this process, we have learned that in order to disentangle different interaction mechanisms it is also necessary to specify the mechanism through which these layers of interactions are combined to yield the observed topology. Within this scenario, block models in different layers do not need to be orthogonal, however, the separability of the layers largely depends on whether they are described by close to orthogonal block models.

In the future, we expect that we will be able to produce a refined and comprehensive method that is able to identify different sets of block models that capture different correlations in the observed connectivity among groups of nodes. The different groupings of nodes corresponding to each block model are supposed to capture different characteristics of the nodes and give an explanation of the mechanisms underlying the creation of a certain network. Because most systems network studies from a global perspective have focused on the identification of a single generative block model (usually called community structure), the possibility of identifying several complementary block structures, it will open a new window toward the understanding of underlying mechanisms responsible for the network's evolution and organization, especially in some biological networks in which the community structure often has not been able to provide new insight.

At the end of the project Marta Sales-Pardo has a permanent position at the host instiution (Universitat Rovirai Virgili) and has started a research lab (1st thesis defended in January 2013). Additionally, in 2013 Marta Sales-Pardo recevied an ICREA Academia award, an award that allows university professors in Catalan universities with an excellent track record to reduce the teaching load and intensify the research tasks.

For more details please contact: marta.sales@urv.cat http://seeslab.info @sees_lab