Social network analysis and Bayesian networks: Taxonomy of environmental contrasts

The main objective of this research is to connect Social Network Analysis descriptive measures of centrality with a probabilistic approach to networks. A latent conditional structure of the network which shapes both the structural information and uncertainty characterizing the data is assumed. This is achieved by means of a Bayesian Network approach which allows reweighting the original adjacency matrix according to the latent network structure. Therefore, traditional SNA measures for network description can be easily computed. This is illustrated using a dataset on Issue Correlates of War (ICOW) which contains claims in world politics for three dimensions of interests, land, river and sea for about 244 claims located in 5 world regions and observed from 1816 to 2001. Each node represents a country involved in single or multiple disputes over the use of natural resources. Thus each dyad is represented by an edge connecting the countries involved in the contrast. Each observed edge is a realization of random variables which are connected by a conditional latent structure. As an example, we consider the duration, the geographical location and the issue of the disputes as main variables characterizing the conflicts. Then, the statistical model provides a pictorial representation of dependence relations between variables by means of modular diagrams, namely Directed Acyclic Graphs (DAGs). The presence of an edge joining nodes in the DAG is interpreted as statistical dependence between the variables. The network is `learned' from data on international conflicts and the arc strengths are used as scores for building a new `observational' adjacency matrix.