This paper introduces a methodology for evaluating network community detection algorithms using ground-truth communities defined by explicit group memberships in 230 large social, collaboration, and information networks. The authors propose a set of 13 structural definitions of network communities and evaluate their performance in identifying ground-truth communities. They find that two definitions, Conductance and Triad-participation-ratio, consistently perform best. The authors also develop a parameter-free community detection method based on local spectral clustering that scales to networks with over 100 million nodes and achieves a 30% relative improvement in F1-score over existing methods. The study highlights the importance of ground-truth in evaluating community detection algorithms and demonstrates that different structural definitions of communities are heavily correlated and naturally group into four classes. The results show that community scoring functions based on triadic closure and conductance best capture the structure of ground-truth communities. The authors also investigate the task of detecting communities from a single seed node and find that their method reliably detects ground-truth communities. The study provides a comprehensive evaluation of community detection methods and highlights the need for more rigorous evaluation standards in the field.This paper introduces a methodology for evaluating network community detection algorithms using ground-truth communities defined by explicit group memberships in 230 large social, collaboration, and information networks. The authors propose a set of 13 structural definitions of network communities and evaluate their performance in identifying ground-truth communities. They find that two definitions, Conductance and Triad-participation-ratio, consistently perform best. The authors also develop a parameter-free community detection method based on local spectral clustering that scales to networks with over 100 million nodes and achieves a 30% relative improvement in F1-score over existing methods. The study highlights the importance of ground-truth in evaluating community detection algorithms and demonstrates that different structural definitions of communities are heavily correlated and naturally group into four classes. The results show that community scoring functions based on triadic closure and conductance best capture the structure of ground-truth communities. The authors also investigate the task of detecting communities from a single seed node and find that their method reliably detects ground-truth communities. The study provides a comprehensive evaluation of community detection methods and highlights the need for more rigorous evaluation standards in the field.