Al.8. However, there are several drawbacks to the GN benchmark: All nodes have the same expected degree, communities are separated in the same way, and the network is of an unrealistic small size. It is a well established fact that most real complex networks are characterised by largely heterogeneous degree distributions1,2,9 and heterogeneous community sizes10?2. For this reason, the GN benchmark cannot be considered as a good proxy for a real network. By consequence, in a newer stream of research5,13, the authors proposed an alternative benchmark, which is usually referred to as LFR (for Lancichinetti, Stattic web Fortunato Radicchi). This method introduces power-law distributions of degree and community size to the graphs to generalise the GN benchmark. The performances of most existing community detection algorithms are good on the GN benchmark. In SIS3 supplier contrast, the LFR benchmark presents a harder test for algorithms and makes it easier to unveil their limitations. It has been shown that the mixing parameter, which is defined as? i k iext i k itot(1)is the most influential parameter in the LFR benchmark graphs stand for the external degree of node i, i.e. the number of edges connecting it to others that belong to different communities, and the total degree of said node. Although it would be possible to define a mixing parameter for each node, it is assumed that is a global property and is the same for every node in the LFR benchmark. The reason here is to be consistent with the standard hypotheses of the planted l-partition model15. According to the definition of community in a strong sense, each node should have more connections within the community than with the rest of the graph16. Therefore, for > 1/2 communities in the strong sense disappear. However, it is worth to mention that Lancichinetti and Fortunato15 found a weaker condition for community detection which can be applied to any version of the planted l-partition model: ?< (N - ncmax )/N , where N is the total number of nodes, and ncmax is the size of the largest community. In our study, although we stick to the strong definition of communities, we have also taken the general condition of into consideration (see Table 1). In the following, we briefly review studies comparing community detection algorithms in chronological order5,8,13?5,17,18 to highlight the research interests shift. In one of the early studies in comparing community detection algorithms, Danon et al. had tested ten algorithms on the GN benchmark78 and collected estimates of how time complexity scales with network observables. However, the authors were not able to compare the actual computational effort as a result of the small sizes of graphs. Later on, Lancichinetti et al. had employed the LFR benchmark to measure the accuracy of two algorithms on undirected unweighted networks without overlapping communities5 and two algorithms on directed weighted networks with overlapping communities13. Concurrently, the authors tested twelve different algorithms on the GN and LFR benchmarks, and random graphs. For the tests on the LFR benchmark, the authors had considered various parameters, including undirected unweighted graphs with non-overlapping communities, directed unweighted graphs with non-overlapping communities, undirected weighted graphs with non-overlapping communities, and undirected unweighted graphs with overlapping communities15. Orman and Labatut later tested five community detection algorithms on the LFR benc.Al.8. However, there are several drawbacks to the GN benchmark: All nodes have the same expected degree, communities are separated in the same way, and the network is of an unrealistic small size. It is a well established fact that most real complex networks are characterised by largely heterogeneous degree distributions1,2,9 and heterogeneous community sizes10?2. For this reason, the GN benchmark cannot be considered as a good proxy for a real network. By consequence, in a newer stream of research5,13, the authors proposed an alternative benchmark, which is usually referred to as LFR (for Lancichinetti, Fortunato Radicchi). This method introduces power-law distributions of degree and community size to the graphs to generalise the GN benchmark. The performances of most existing community detection algorithms are good on the GN benchmark. In contrast, the LFR benchmark presents a harder test for algorithms and makes it easier to unveil their limitations. It has been shown that the mixing parameter, which is defined as? i k iext i k itot(1)is the most influential parameter in the LFR benchmark graphs stand for the external degree of node i, i.e. the number of edges connecting it to others that belong to different communities, and the total degree of said node. Although it would be possible to define a mixing parameter for each node, it is assumed that is a global property and is the same for every node in the LFR benchmark. The reason here is to be consistent with the standard hypotheses of the planted l-partition model15. According to the definition of community in a strong sense, each node should have more connections within the community than with the rest of the graph16. Therefore, for > 1/2 communities in the strong sense disappear. However, it is worth to mention that Lancichinetti and Fortunato15 found a weaker condition for community detection which can be applied to any version of the planted l-partition model: ?< (N - ncmax )/N , where N is the total number of nodes, and ncmax is the size of the largest community. In our study, although we stick to the strong definition of communities, we have also taken the general condition of into consideration (see Table 1). In the following, we briefly review studies comparing community detection algorithms in chronological order5,8,13?5,17,18 to highlight the research interests shift. In one of the early studies in comparing community detection algorithms, Danon et al. had tested ten algorithms on the GN benchmark78 and collected estimates of how time complexity scales with network observables. However, the authors were not able to compare the actual computational effort as a result of the small sizes of graphs. Later on, Lancichinetti et al. had employed the LFR benchmark to measure the accuracy of two algorithms on undirected unweighted networks without overlapping communities5 and two algorithms on directed weighted networks with overlapping communities13. Concurrently, the authors tested twelve different algorithms on the GN and LFR benchmarks, and random graphs. For the tests on the LFR benchmark, the authors had considered various parameters, including undirected unweighted graphs with non-overlapping communities, directed unweighted graphs with non-overlapping communities, undirected weighted graphs with non-overlapping communities, and undirected unweighted graphs with overlapping communities15. Orman and Labatut later tested five community detection algorithms on the LFR benc.
