Abstract: Networks are pervasive for complex systems modeling, from biology tolanguage or social sciences, ecosystems or computer science. Detecting com-munities in networks is among the main methods to reveal meaningful struc-tural patterns for the understanding of those systems. Although dozens ofclustering methods have been proposed so far, sometimes including parame-ters such as resolution or scaling, there is no unified framework for selectingthe method best suited to a research objective. After more than 20 years ofresearch, scientists still justify their methodological choice based on ad-hoccomparisons with ‘ground-truth’ or synthetic networks, making it challengingto perform comparative study between those methods. This paper proposesa unified framework, based on easy-to-understand measures, that enables theselection of appropriate clustering methods according to the situation. If re-quired, it can also be used to fine-tune their parameters by interpreting themas description scale parameters. We demonstrate that a new family of algo-rithms inspired by our approach outperforms a set of state-of-the-art com-munity detection algorithms, by comparing them on a benchmark dataset.We believe our approach has the potential to provide a fresh start and a solidfoundation for the development and evaluation of clustering methods acrossa wide range of disciplines.
No Comments.