Herefore not reliable. Thus, the value of ?must be recalculated to get an idea of the repeatability of the results and confirm its validity. In some situations, one might need to repeat the detection processes several times or switch to another algorithm to ensure the validity of the community detection results.Scientific RepoRts | 6:30750 | DOI: 10.1038/srepwww.nature.com/XAV-939 site scientificreports/Fastgreedy R2 R2 2.048 [0.006] 0.956 Multilevel 1.126 [0.003] 0.957 Infomap 1.421 [0.009] 0.933 Walktrap 2.04 [0.002] 0.962 Leading eigenvector 1.123 [0.005] 0.951 Spinglass 1.282 [0.013] () 0.867 Label propagation 0.959 [0.005] 0.947 Edge betweenness 2.915 [0.005] 0.Table 2. Indexes of the exponential function T N with the corresponding adjusted Stattic custom synthesis R-squared values. The standard errors are listed in brackets. All the results are statistically significant at the significance level of 0.05. Spinglass and Edge betweenness algorithms have been tested only on small networks with N 1000, there might be some biases in the indexes of these two methods.Figure 7. Recommendation for the choice of adaptable community detection algorithms. The x-axis is the mixing parameter and the y-axis is the number of nodes N. The y-axis is on a log scale for better visualisation. The coordinates of certain important points are: A(0.48, 1000), B(0.6, 1000), C(0.48, 6192), D(0.36, 31948), and E(0.42, 31948). In different regions we would like to recommend different algorithms, which are represented by different abbreviations: IM is the Infomap algorithm, LP is the Label propagation algorithm, ML is the Multilevel algorithm, WT is the Walktrap algorithm, SG is the Spinglass algorithm, and EB represents the Edge betweenness algorithm.Our suggestions have to be applied in conjunction with the concomitant research questions. As a pure application of the recommendations could bias the results. Once a researcher has decided to use a specific community detection algorithm, it is of crucial importance for her to keep in mind the limitations and the expected validity of the output of the community detection algorithm chosen. It is noteworthy that metadata would be helpful for evaluating network community detection methods and can be used to improve the analysis and understanding of network structure19,27. In real-world networks where metadata is available, researchers should also take into account the research question, the properties of the network, the interpretation and meaning of the communities while choosing the community detection algorithms. Different research questions together with the metadata might lead to different definitions of community, and further change the ground truth of the network. Compared to previous works on benchmarking community detection algorithms, our study has many obvious advantages: First, we have considered networks which contain a wide spectrum of number of nodes and mixing parameters. Second, the algorithms we have tested are integrated in a cross-platform package which has been widely used in academic research in network science and related fields. Third, we have used the LFR benchmark graphs which have shown more realistic properties than the earlier computer-generated networks such as the GN benchmark.Scientific RepoRts | 6:30750 | DOI: 10.1038/srepwww.nature.com/scientificreports/Figure 8. Suggestion for the community detection process. Small networks are those with number of nodes less than 1000, and small corresponds to ?0.5. To be noticed.Herefore not reliable. Thus, the value of ?must be recalculated to get an idea of the repeatability of the results and confirm its validity. In some situations, one might need to repeat the detection processes several times or switch to another algorithm to ensure the validity of the community detection results.Scientific RepoRts | 6:30750 | DOI: 10.1038/srepwww.nature.com/scientificreports/Fastgreedy R2 R2 2.048 [0.006] 0.956 Multilevel 1.126 [0.003] 0.957 Infomap 1.421 [0.009] 0.933 Walktrap 2.04 [0.002] 0.962 Leading eigenvector 1.123 [0.005] 0.951 Spinglass 1.282 [0.013] () 0.867 Label propagation 0.959 [0.005] 0.947 Edge betweenness 2.915 [0.005] 0.Table 2. Indexes of the exponential function T N with the corresponding adjusted R-squared values. The standard errors are listed in brackets. All the results are statistically significant at the significance level of 0.05. Spinglass and Edge betweenness algorithms have been tested only on small networks with N 1000, there might be some biases in the indexes of these two methods.Figure 7. Recommendation for the choice of adaptable community detection algorithms. The x-axis is the mixing parameter and the y-axis is the number of nodes N. The y-axis is on a log scale for better visualisation. The coordinates of certain important points are: A(0.48, 1000), B(0.6, 1000), C(0.48, 6192), D(0.36, 31948), and E(0.42, 31948). In different regions we would like to recommend different algorithms, which are represented by different abbreviations: IM is the Infomap algorithm, LP is the Label propagation algorithm, ML is the Multilevel algorithm, WT is the Walktrap algorithm, SG is the Spinglass algorithm, and EB represents the Edge betweenness algorithm.Our suggestions have to be applied in conjunction with the concomitant research questions. As a pure application of the recommendations could bias the results. Once a researcher has decided to use a specific community detection algorithm, it is of crucial importance for her to keep in mind the limitations and the expected validity of the output of the community detection algorithm chosen. It is noteworthy that metadata would be helpful for evaluating network community detection methods and can be used to improve the analysis and understanding of network structure19,27. In real-world networks where metadata is available, researchers should also take into account the research question, the properties of the network, the interpretation and meaning of the communities while choosing the community detection algorithms. Different research questions together with the metadata might lead to different definitions of community, and further change the ground truth of the network. Compared to previous works on benchmarking community detection algorithms, our study has many obvious advantages: First, we have considered networks which contain a wide spectrum of number of nodes and mixing parameters. Second, the algorithms we have tested are integrated in a cross-platform package which has been widely used in academic research in network science and related fields. Third, we have used the LFR benchmark graphs which have shown more realistic properties than the earlier computer-generated networks such as the GN benchmark.Scientific RepoRts | 6:30750 | DOI: 10.1038/srepwww.nature.com/scientificreports/Figure 8. Suggestion for the community detection process. Small networks are those with number of nodes less than 1000, and small corresponds to ?0.5. To be noticed.