The impact of interaction and algorithm choice on identified communities

In social networks, nodes are organized into densely linked communities where edges appear among the nodes with high concentration. Identifying communities has proven to be a challenging task due to various community definitions/algorithms and also due to the lack of “ground truth” for reference and evaluation. These communities not only differ due to various definitions but also can be affected by the type of interactions modeled in the network, which lead to different social groups. We are interested in exploring and studying the concept of partial network views, which is based on multiple types of interactions. An Enron email network is used to conduct our experiments. In this paper, we explore the mutual impact of selecting different views extracted from the same network and their interplay with various community detection algorithms to measure the change and the level of realism of the structure for non-overlapping communities. To better understand this, we assess the agreement of partitions by evaluating the partitioning quality (performance) and finding the similarity between algorithms. The results demonstrate that the topological properties of communities and the performance of algorithms are equivalent to each other. Both of them are affected by the type of interaction specified in each view. Some network views appeared to have more interesting communities than other views, thus, might help to approach a relatively informative and logic “ground truth” for communities.