def test_fair_personalizer():
H = pg.PageRank(assume_immutability=True, normalization="symmetric")
algorithms = {
"FairPers":
lambda G, p, s: pg.Normalize(
pg.FairPersonalizer(H, error_type=pg.Mabs, max_residual=0)).rank(
G, p, sensitive=s),
"FairPers-C":
lambda G, p, s: pg.Normalize(
pg.FairPersonalizer(
H, .80, pRule_weight=10, error_type=pg.Mabs, max_residual=0)).
rank(G, p, sensitive=s),
"FairPersSkew":
lambda G, p, s: pg.Normalize(
pg.FairPersonalizer(H, error_skewing=True, max_residual=0)).rank(
G, p, sensitive=s),
"FairPersSkew-C":
lambda G, p, s: pg.Normalize(
pg.FairPersonalizer(
H, .80, error_skewing=True, pRule_weight=10, max_residual=0)
).rank(G, p, sensitive=s),
}
_, graph, groups = next(pg.load_datasets_multiple_communities(["bigraph"]))
labels = pg.to_signal(graph, groups[0])
sensitive = pg.to_signal(graph, groups[1])
for algorithm in algorithms.values():
ranks = algorithm(graph, labels, sensitive)
assert pg.pRule(sensitive)(
ranks
) > 0.79 # allow a leeway for generalization capabilities compared to 80%
def test_explicit_citations():
assert "unknown node ranking algorithm" == pg.NodeRanking().cite()
assert "with parameters tuned \cite{krasanakis2021pygrank}" in pg.ParameterTuner(
lambda params: pg.PageRank(params[0])).cite()
assert "Postprocessor" in pg.Postprocessor().cite()
assert pg.PageRank().cite() in pg.AlgorithmSelection().cite()
assert "krasanakis2021pygrank" in pg.ParameterTuner().cite()
assert "ortega2018graph" in pg.ParameterTuner().cite()
assert pg.HeatKernel().cite() in pg.SeedOversampling(pg.HeatKernel()).cite()
assert pg.AbsorbingWalks().cite() in pg.BoostedSeedOversampling(pg.AbsorbingWalks()).cite()
assert "krasanakis2018venuerank" in pg.BiasedKernel(converge_to_eigenvectors=True).cite()
assert "yu2021chebyshev" in pg.HeatKernel(coefficient_type="chebyshev").cite()
assert "susnjara2015accelerated" in pg.HeatKernel(krylov_dims=5).cite()
assert "krasanakis2021pygrank" in pg.GenericGraphFilter(optimization_dict=dict()).cite()
assert "tautology" in pg.Tautology().cite()
assert pg.PageRank().cite() == pg.Tautology(pg.PageRank()).cite()
assert "mabs" in pg.MabsMaintain(pg.PageRank()).cite()
assert "max normalization" in pg.Normalize(pg.PageRank()).cite()
assert "[0,1] range" in pg.Normalize(pg.PageRank(), "range").cite()
assert "ordinal" in pg.Ordinals(pg.PageRank()).cite()
assert "exp" in pg.Transformer(pg.PageRank()).cite()
assert "0.5" in pg.Threshold(pg.PageRank(), 0.5).cite()
assert "andersen2007local" in pg.Sweep(pg.PageRank()).cite()
assert pg.HeatKernel().cite() in pg.Sweep(pg.PageRank(), pg.HeatKernel()).cite()
assert "LFPRO" in pg.AdHocFairness("O").cite()
assert "LFPRO" in pg.AdHocFairness(pg.PageRank(), "LFPRO").cite()
assert "multiplicative" in pg.AdHocFairness(pg.PageRank(), "B").cite()
assert "multiplicative" in pg.AdHocFairness(pg.PageRank(), "mult").cite()
assert "tsioutsiouliklis2020fairness" in pg.AdHocFairness().cite()
assert "rahman2019fairwalk" in pg.FairWalk(pg.PageRank()).cite()
assert "krasanakis2020prioredit" in pg.FairPersonalizer(pg.PageRank()).cite()
def test_invalid_fairness_arguments():
_, graph, groups = next(pg.load_datasets_multiple_communities(["bigraph"]))
labels = pg.to_signal(graph, groups[0])
sensitive = pg.to_signal(graph, groups[1])
H = pg.PageRank(assume_immutability=True, normalization="symmetric")
with pytest.raises(Exception):
# this tests that a deprecated way of applying fairwalk actually raises an exception
pg.AdHocFairness(H, method="FairWalk").rank(graph, labels, sensitive=sensitive)
with pytest.raises(Exception):
pg.FairPersonalizer(H, parity_type="universal").rank(graph, labels, sensitive=sensitive)
with pytest.raises(Exception):
pg.FairWalk(None).transform(H.rank(graph, labels), sensitive=sensitive)
def test_fair_personalizer_mistreatment():
H = pg.PageRank(assume_immutability=True, normalization="symmetric")
algorithms = {
"Base": lambda G, p, s: H.rank(G, p),
"FairPersMistreat": pg.Normalize(pg.FairPersonalizer(H, parity_type="mistreatment", pRule_weight=10)),
"FairPersTPR": pg.Normalize(pg.FairPersonalizer(H, parity_type="TPR", pRule_weight=10)),
"FairPersTNR": pg.Normalize(pg.FairPersonalizer(H, parity_type="TNR", pRule_weight=-1)) # TNR optimization increases mistreatment for this example
}
mistreatment = lambda known_scores, sensitive_signal, exclude: \
pg.AM([pg.Disparity([pg.TPR(known_scores, exclude=1 - (1 - exclude) * sensitive_signal),
pg.TPR(known_scores, exclude=1 - (1 - exclude) * (1 - sensitive_signal))]),
pg.Disparity([pg.TNR(known_scores, exclude=1 - (1 - exclude) * sensitive_signal),
pg.TNR(known_scores, exclude=1 - (1 - exclude) * (1 - sensitive_signal))])])
_, graph, groups = next(pg.load_datasets_multiple_communities(["synthfeats"]))
labels = pg.to_signal(graph, groups[0])
sensitive = pg.to_signal(graph, groups[1])
train, test = pg.split(labels)
# TODO: maybe try to check for greater improvement
base_mistreatment = mistreatment(test, sensitive, train)(algorithms["Base"](graph, train, sensitive))
for algorithm in algorithms.values():
if algorithm != algorithms["Base"]:
print(algorithm.cite())
assert base_mistreatment >= mistreatment(test, sensitive, train)(algorithm(graph, train, sensitive))
for name, filter in filters.items():
print("=====", name, "=====")
algorithms = {
"None":
filter,
"Mult":
pg.AdHocFairness(filter, "B"),
"LFPRO":
pg.AdHocFairness(filter, "O"),
#"FBuck-C": pg.FairPersonalizer(filter, .8, pRule_weight=10, max_residual=1, error_type=pg.Mabs, parameter_buckets=0),
"FPers-C":
pg.FairPersonalizer(filter,
.8,
pRule_weight=10,
max_residual=0,
error_type=pg.Mabs,
error_skewing=True,
parity_type="impact"),
"Fest-C":
pg.FairPersonalizer(filter,
.8,
pRule_weight=10,
max_residual=1,
error_type=pg.Mabs,
error_skewing=False,
parameter_buckets=1,
parity_type="impact")
#"FFfix-C": pg.FairTradeoff(filter, .8, pRule_weight=10, error_type=pg.Mabs)
#"FairTf": pg.FairnessTf(filter)
}