Potential Parameters

• CommunityImputer
  • How to use the community information for imputation
    • Use the feature's mean within the community
    • Use the feature's median within the community
      • probably only useful for non-continuous features
    • Use a weighted mean based on some community-intrinsic centrality or similarity measure
      • It is much cheaper to calculate these values in the limited scopes of our pre-computed communities
  • What to do if feature is not observed within the community
    • Use values of neighboring communities
    • Use global average
    • Use some constant
• Louvain method (probably not that important)
  • max_levels
  • max_iterations
  • tolerance / min_modularity_gain
  • resolution parameter to adjust the size of communities
    • see implementation in NetworkX for example

Even more ideas (probably no time left)

• Run Louvain multiple times and, if a node is in multiple communities, use the multiple community averages to impute the node average

Formulas

• Formula for calculating modularity gain from moving a node:
  • 
  • Source: Implementation by the authors of the Louvain method
• Formula for calculating graph modularity:
  • 
  • Source