Data-driven analysis of complex networks has been in the focus of research for decades. An important area of research is to study how well real networks can be described with a small selection of metrics, furthermore how well network models can capture the relations between graph metrics observed in real networks. In this paper, we apply machine-learning techniques to investigate the aforementioned problems. We study 500 real-world networks along with 2000 synthetic networks generated by four frequently used network models with previously calibrated parameters to make the generated graphs as similar to the real networks as possible. This paper unifies several branches of data-driven complex network analysis, such as the study of graph metrics and their pair-wise relationships, network similarity estimation, model calibration, and graph classification. We find that the correlation profiles of the structural measures significantly differ across network domains and the domain can be efficiently determined using a small selection of graph metrics. The structural properties of the network models with fixed parameters are robust enough to perform parameter calibration. The goodness-of-fit of the network models highly depends on the network domain. By solving classification problems, we find that the models lack the capability of generating a graph with a high clustering coefficient and relatively large diameter simultaneously. On the other hand, models are able to capture exactly the degree-distribution-related metrics.

Network representations of systems from various scientific and societal domains are neither completely random nor fully regular, but instead appear to contain recurring structural features. These features tend to be shared by networks belonging to the same broad class, such as the class of social networks or the class of biological networks. Within each such class, networks describing similar systems tend to have similar features. This occurs presumably because networks representing similar systems would be expected to be generated by a shared set of domain-specific mechanisms, and it should therefore be possible to classify networks based on their features at various structural levels. Here we describe and demonstrate a new hybrid approach that combines manual selection of network features of potential interest with existing automated classification methods. In particular, selecting well-known network features that have been studied extensively in social network analysis and network science literature, and then classifying networks on the basis of these features using methods such as random forest, which is known to handle the type of feature collinearity that arises in this setting, we find that our approach is able to achieve both higher accuracy and greater interpretability in shorter computation time than other methods.