Secondary abstract: |
When researching relationships between data entities, the most natural way of presenting them is by using networks. When constructing networks from data, the lack of relevant data often prevents us from building a complete network. In such cases, we are only able to build small or incomplete networks, which are of very limited use in the further analysis. We then often solve this problem by constructing new, random networks.
This paper presents a new approach to generating random networks, which is called M-generator. The task of M-generator is to automatically analyze the available network, and on the basis of selected properties generate a random network that follows these properties. To select optimal network model to generate random network, we use a machine learning method, based on the analysis of the original network. Analysis and selection of a random network is fully automated, so that the presence of a domain expert in selecting network properties and selecting a random network model is not required.
Model operation was tested on real world data, where random network properties seemed to follow the real world ones. However, due to slightly smaller sample size and the lack of labelled data, we can not estimate the efficiency in general. Despite that, we are satisfied with the results, as we managed to automatically generate really good random networks. |