This paper addresses the challenge of identifying communities in real-world networks, which are often densely linked and form natural groups. The authors study 230 large social, collaboration, and information networks where nodes explicitly state their group memberships, providing a reliable ground-truth for evaluating community detection methods. They propose a methodology to compare different structural definitions of network communities with the ground-truth communities, focusing on their sensitivity, robustness, and performance. The study reveals that 13 commonly used structural definitions naturally group into four classes, and two definitions—Conductance and Triad-participation-ratio—perform consistently well. Additionally, the authors extend the local spectral clustering algorithm to a parameter-free community detection method that scales to networks with hundreds of millions of nodes, achieving a 30% relative improvement over current methods. The paper also explores the task of detecting communities from a single seed node, demonstrating the effectiveness of the proposed method. The availability of ground-truth communities and the datasets used in this study will contribute to more rigorous evaluation and future research in network community detection.This paper addresses the challenge of identifying communities in real-world networks, which are often densely linked and form natural groups. The authors study 230 large social, collaboration, and information networks where nodes explicitly state their group memberships, providing a reliable ground-truth for evaluating community detection methods. They propose a methodology to compare different structural definitions of network communities with the ground-truth communities, focusing on their sensitivity, robustness, and performance. The study reveals that 13 commonly used structural definitions naturally group into four classes, and two definitions—Conductance and Triad-participation-ratio—perform consistently well. Additionally, the authors extend the local spectral clustering algorithm to a parameter-free community detection method that scales to networks with hundreds of millions of nodes, achieving a 30% relative improvement over current methods. The paper also explores the task of detecting communities from a single seed node, demonstrating the effectiveness of the proposed method. The availability of ground-truth communities and the datasets used in this study will contribute to more rigorous evaluation and future research in network community detection.