• Neighborhood: k-hop or k-nearest
• Size of the neighborhood $k$
• How to use the observed features in the neighborhood for imputation
  • Mean
  • Median (only really useful for non-continuous features)
  • Weighted mean (e.g. weighted by the distance to the source node)
• Whether to use imputed features for future imputations
  • If not: flag imputed features, so that they may be ignored for future imputations
• What to do if feature is not present in neighborhood
  • Use global average
  • Continue searching until the feature is found (or until it was found a certain number of times)
  • Use constant value

• Strategy for filling features that couldn't be imputed from the neighbourhood
  • Before: calculate global average of the feature on-demand and every time this case occurs
  • Current: calculate global average of the feature on-demand and store the result (e.g. in a HashMap), instead of calculating it repeatedly
    • might lead to worse quality, because changes in the actual global average will not be represented
    • will increase solving times, especially in graphs with a lot of isolated nodes
• Parallelization
• Use iterators instead of copying vectors (e.g. in get_neighbours)
• In the second for-loop: Iterate over neighbourhood instead of iteration over the features (see draft)
  • Might improve cache efficiency because all operations on the feature vector of a neighbour are performed in a batch (see principle of locality).
  • Will need three additional arrays, on the other hand.
  • → Benchmark this

Initial draft with explanation for design decisions