[Paper Review] Link community detection through global optimization and the inverse resolution limit of partition density

Juyong Lee, Zhongyuan Zhang+3 more|arXiv (Cornell University)|Jan 19, 2016

Complex Network Analysis Techniques41 references3 citations

TL;DR

This paper identifies a critical flaw in partition density—the quality function used in link community detection—showing it exhibits an 'inverse resolution limit' that strongly favors small cliques, especially triangles, over larger communities. To overcome this, the authors propose a novel weighted line graph transformation using a normalized Jaccard index, which, when optimized via modularity or Infomap, yields significantly more accurate and meaningful link communities than prior methods.

ABSTRACT

We investigate the possibility of global optimization-based overlapping community detection, using link community framework. We first show that partition density, the original quality function used in link community detection method, is not suitable as a quality function for global optimization because it prefers breaking communities into triangles except in highly limited conditions. We analytically derive those conditions and confirm it with computational results on direct optimization of various synthetic and real-world networks. To overcome this limitation, we propose alternative approaches combining the weighted line graph transformation and existing quality functions for node-based communities. We suggest a new line graph weighting scheme, a normalized Jaccard index. Computational results show that community detection using the weighted line graphs generated with the normalized Jaccard index leads to a more accurate community structure.

Motivation & Objective

• To investigate whether partition density is suitable as a quality function for global optimization in link community detection.
• To identify the root cause of partition density's preference for small cliques, particularly triangles, in network partitioning.
• To develop a robust alternative method for link community detection that avoids the inverse resolution limit of partition density.
• To evaluate the performance of weighted line graph transformations using different similarity metrics in detecting accurate community structures.
• To propose and validate a new line graph weighting scheme—normalized Jaccard index—that enhances community detection accuracy via global optimization.

Proposed method

• The authors analytically derive the conditions under which partition density favors triangle formation, showing that the local partition density of a triangle is always 1.0, which is often higher than the reduced density of larger communities after splitting.
• They introduce a new line graph weighting scheme based on the normalized Jaccard index to better reflect link similarity in the original network, combining topological and structural information.
• The method transforms the original network into a weighted line graph where edge weights represent the normalized Jaccard similarity between links.
• Disjoint node communities in the line graph are detected using standard global optimization techniques such as modularity maximization and the Infomap algorithm.
• The performance of the proposed method is evaluated on synthetic LFR benchmark networks with overlapping communities and real-world networks, using normalized mutual information (NMI) as the evaluation metric.
• The approach is compared against two baseline weighting schemes: simple normalization and standard Jaccard index, to isolate the impact of normalization on performance.

Experimental results

Research questions

• RQ1 Does partition density suffer from a systematic bias toward detecting small cliques, particularly triangles, in link community detection?
• RQ2 Under what conditions does partition density fail to detect larger, more meaningful communities due to this bias?
• RQ3 Can a weighted line graph transformation with a normalized Jaccard index mitigate the inverse resolution limit of partition density?
• RQ4 How does the performance of community detection using the normalized Jaccard index compare to other weighting schemes in terms of accuracy on synthetic and real-world networks?
• RQ5 Is global optimization of modularity on the transformed line graph more effective than hierarchical clustering based on partition density for link community detection?

Key findings

• Partition density exhibits a strong preference for triangles due to their maximal local partition density of 1.0, which often outweighs the marginal decrease in density from splitting larger communities.
• The only exception where a triangle is not preferred is when four nodes are connected by five links, indicating a rare scenario where larger communities remain stable.
• Direct global optimization of partition density on synthetic and real-world networks results in a high number of spurious triangle communities, indicating poor performance.
• The normalized Jaccard index-based line graph weighting scheme significantly improves community detection accuracy, as evidenced by higher normalized mutual information (NMI) scores on LFR benchmark networks.
• Optimization of modularity on line graphs generated with the normalized Jaccard index outperforms both the standard Jaccard and simple normalization schemes across varying levels of network overlap and mixing parameters.
• The proposed method effectively mitigates the inverse resolution limit of partition density, enabling detection of larger, more meaningful link communities in complex networks.