I use the "Louvain" algorithm (Blondel et al., 2008) in Gephi to detect communities in graphs. This algorithm provides a modularity value Q as sort a "quality-control". The higher the value is, the best the result should be.
Gephi (versions < 0.7) said that it Q < 0.4, the result should not be considered too seriously. Is this an absolute answer? Are partitions with Q < 0.4 really bad results? Uninterpretable results? Nonsense results?
Thank you for your answer.
Your example with the Zachary Karate club raises another question : do the modularity aim at evaluating groups that already exist ? I think it only can tell if a algorithm is efficient or not. But the communities you refer to (I think) are "real-life" people groups, that aim not at maximizing a Q value. Groups people made spontaneously don't aim (I think) at having the nice repartition evaluated by Q.
Do you know if there is any Gephi plugin (or another tool with GUI) that implements other methods of graph segmentation ?
Hi, didn't know the karate club dataset had such a low modularity, but never checked either. Interesting! I think the point is that you want to find "real" groups, and the question is how good modularity is a measure of success. If you already *know* the real groups, and modularity isn't high on that grouping, it's a bad sign for modularity, and in particular that being below 0.4 does not mean that you haven't found the "real" structure you're looking for. I think the maximally obtainable modularity in a graph is a function of the graph's connectivity, so I'd rather compare a solution's value to the value of other possible solutions.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Graphs