def test_all_communities_benchmarks(): datasets = ["bigraph"] pre = pg.preprocessor(assume_immutability=True, normalization="symmetric") tol = 1.E-9 optimization = pg.SelfClearDict() algorithms = { "ppr0.85": pg.PageRank(alpha=0.85, preprocessor=pre, max_iters=10000, tol=tol), "ppr0.9": pg.PageRank(alpha=0.9, preprocessor=pre, max_iters=10000, tol=tol), "ppr0.99": pg.PageRank(alpha=0.99, preprocessor=pre, max_iters=10000, tol=tol), "hk3": pg.HeatKernel(t=3, preprocessor=pre, max_iters=10000, tol=tol, optimization_dict=optimization), "hk5": pg.HeatKernel(t=5, preprocessor=pre, max_iters=10000, tol=tol, optimization_dict=optimization), "hk7": pg.HeatKernel(t=7, preprocessor=pre, max_iters=10000, tol=tol, optimization_dict=optimization), } tuned = {"selected": pg.AlgorithmSelection(algorithms.values(), fraction_of_training=0.8)} loader = pg.load_datasets_all_communities(datasets, min_group_size=50) pg.benchmark_print(pg.benchmark(algorithms | tuned, loader, pg.AUC, fraction_of_training=.8, seed=list(range(1))), decimals=3, delimiter=" & ", end_line="\\\\") loader = pg.load_datasets_all_communities(datasets, min_group_size=50) pg.benchmark_print(pg.benchmark(algorithms | tuned, loader, pg.Modularity, sensitive=pg.pRule, fraction_of_training=.8, seed=list(range(1))), decimals=3, delimiter=" & ", end_line="\\\\") mistreatment = lambda known_scores, sensitive_signal, exclude: \ pg.AM([pg.Disparity([pg.TPR(known_scores, exclude=1 - (1 - exclude.np) * sensitive_signal.np), pg.TPR(known_scores, exclude=1 - (1 - exclude.np) * (1 - sensitive_signal.np))]), pg.Disparity([pg.TNR(known_scores, exclude=1 - (1 - exclude.np) * sensitive_signal.np), pg.TNR(known_scores, exclude=1 - (1 - exclude.np) * (1 - sensitive_signal.np))])]) loader = pg.load_datasets_all_communities(datasets, min_group_size=50) pg.benchmark_print(pg.benchmark(algorithms | tuned, loader, pg.Modularity, sensitive=mistreatment, fraction_of_training=.8, seed=list(range(1))), decimals=3, delimiter=" & ", end_line="\\\\")
def test_best_direction(): assert pg.Conductance().best_direction() == -1 assert pg.Density().best_direction() == 1 assert pg.Modularity().best_direction() == 1 assert pg.AUC([1, 2, 3]).best_direction() == 1 assert pg.Cos([1, 2, 3]).best_direction() == 1 assert pg.Dot([1, 2, 3]).best_direction() == 1 assert pg.TPR([1, 2, 3]).best_direction() == 1 assert pg.TNR([1, 2, 3]).best_direction() == 1
def test_computations(): for _ in supported_backends(): assert pg.Accuracy([1, 2, 3])([1, 2, 3]) == 1 assert pg.Mabs([3, 1, 1])([2, 0, 2]) == 1 assert pg.CrossEntropy([1, 1, 1])([1, 1, 1]) < 1.E-12 assert float(pg.Cos([2, 0, 1])([2, 0, 1])) == 1 assert float(pg.Cos([2, 0, 1])([-2, 0, -1])) == -1 assert float(pg.Cos([0, 0, 0])([0, 0, 0])) == 0 assert float(pg.Dot([1, 1, 1])([1, 1, 1])) == 3 assert float(pg.TPR([1, 0, 0, 0])([1, 1, 0, 0])) == 0.5 assert float(pg.TNR([0, 0, 0, 1])([1, 1, 0, 0])) == 0.5
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))
.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) } algorithms = pg.create_variations(algorithms, {"": pg.Normalize}) #import cProfile as profile #pr = profile.Profile() #pr.enable() mistreatment = lambda known_scores, sensitive_signal, exclude: \ pg.AM([pg.Disparity([pg.TPR(known_scores, exclude=1-(1-exclude.np)*sensitive_signal.np), pg.TPR(known_scores, exclude=1-(1-exclude.np)*(1-sensitive_signal.np))]), pg.Disparity([pg.TNR(known_scores, exclude=1 - (1 - exclude.np) * sensitive_signal.np), pg.TNR(known_scores, exclude=1 - (1 - exclude.np) * (1 - sensitive_signal.np))])]) pg.benchmark_print(pg.benchmark(algorithms, pg.load_datasets_multiple_communities( datasets, max_group_number=2), metric=pg.AUC, sensitive=pg.pRule, fraction_of_training=seed_fractions), delimiter=" & ", end_line="\\\\") #pr.disable() #pr.dump_stats('profile.pstat')