Adjusted homophily and label informativeness

[OlegPlatonov]

Hi! In a recent paper Characterizing Graph Datasets for Node Classification: Homophily-Heterophily Dichotomy and Beyond several characteristics of labeled graphs have been proposed. First, it has been shown that a rarely used in graph ML literature measure named adjusted homophily satisfies more desirable properties than other homophily measures, which makes it appropriate for comparing homophily levels across datasets with different number of classes, class sizes, and degree distributions across classes. Further, a new characteristic - label informativeness - has been proposed. It shows how much information about a node's label we get from knowing its neighbor's label. It has been shown that label informativeness is much more correlated with GNN performance than homophily.

In this PR I've added adjusted homophily and label informativeness to PyG.

[codecov[bot]]

Codecov Report

All modified and coverable lines are covered by tests :white_check_mark:

Comparison is base (d71bab2) 88.73% compared to head (877cd4c) 88.76%. Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #8308      +/-   ##
==========================================
+ Coverage   88.73%   88.76%   +0.02%     
==========================================
  Files         479      480       +1     
  Lines       29820    29903      +83     
==========================================
+ Hits        26462    26543      +81     
- Misses       3358     3360       +2     

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.