Fairness

Fairness-aware graph filters and postprocessors are provided. Below we provide two example node ranking algorithms that make personalized PageRank fairer by supplying it with a postprocessor or straight up using a variation.

import pygrank as pg

ppro = pg.PageRank(alpha=0.9, max_iters=1000, tol=1.E-9) >> pg.AdHocFairness("O")
lfpr = pg.LFPR(alpha=0.9, max_iters=1000, tol=1.E-9)

When calling fairness-aware methods, add a sensitive keyword argument that holds information that together with a graph can construct a graph signal.

Several measures also provide fairness assessment.

_, graph, groups = next(pg.load_datasets_multiple_communities(['citeseer'], max_group_number=2, directed=False))
measure = pg.pRule(groups[1])

pg.benchmark_print_line("ppro", ppro(graph, groups[0], sensitive=groups[1]) >> measure)
pg.benchmark_print_line("lfpr", lfpr(graph, groups[0], sensitive=groups[1]) >> measure)