def test_HawkesExpKern_solver_step(self):
"""...Test HawkesExpKern setting of step parameter
of solver
"""
for solver in solvers:
if solver in ['bfgs']:
msg = '^Solver "%s" has no settable step$' % solver
with self.assertWarnsRegex(RuntimeWarning, msg):
learner = HawkesExpKern(
self.decays,
solver=solver,
step=1,
**Test.specific_solver_kwargs(solver))
self.assertIsNone(learner.step)
else:
learner = HawkesExpKern(self.decays,
solver=solver,
step=self.float_1,
**Test.specific_solver_kwargs(solver))
self.assertEqual(learner.step, self.float_1)
self.assertEqual(learner._solver_obj.step, self.float_1)
learner.step = self.float_2
self.assertEqual(learner.step, self.float_2)
self.assertEqual(learner._solver_obj.step, self.float_2)
if solver in ['sgd']:
msg = '^SGD step needs to be tuned manually$'
with self.assertWarnsRegex(RuntimeWarning, msg):
learner = HawkesExpKern(self.decays,
solver='sgd',
max_iter=1)
learner.fit(self.events, 0.3)
def test_HawkesExpKern_fit(self):
"""...Test HawkesExpKern fit with different solvers
and penalties
"""
sto_seed = 179312
n_nodes = 2
events, baseline, adjacency = Test.get_train_data(n_nodes=n_nodes,
betas=self.decays)
start = 0.3
initial_adjacency_error = \
Test.estimation_error(start * np.ones((n_nodes, n_nodes)),
adjacency)
for gofit in gofits:
for penalty in penalties:
for solver in solvers:
solver_kwargs = {'penalty': penalty, 'tol': 1e-10,
'solver': solver, 'verbose': False,
'max_iter': 10, 'gofit': gofit}
if penalty != 'none':
solver_kwargs['C'] = 50
if solver in ['sgd', 'svrg']:
solver_kwargs['random_state'] = sto_seed
# manually set step
if solver == 'sgd' and gofit == 'likelihood':
solver_kwargs['step'] = 3e-1
elif solver == 'sgd' and gofit == 'least-squares':
solver_kwargs['step'] = 1e-5
elif solver == 'svrg' and gofit == 'likelihood':
solver_kwargs['step'] = 1e-3
elif solver == 'svrg' and gofit == 'least-squares':
continue
if solver == 'bfgs':
# BFGS only accepts ProxZero and ProxL2sq for now
if penalty != 'l2':
continue
if penalty == 'nuclear':
# Nuclear penalty only compatible with batch solvers
if solver in \
HawkesExpKern._solvers_stochastic:
continue
learner = HawkesExpKern(self.decays,
**solver_kwargs)
learner.fit(events, start=start)
adjacency_error = Test.estimation_error(
learner.adjacency, adjacency)
self.assertLess(adjacency_error,
initial_adjacency_error * 0.8,
"solver %s with penalty %s and "
"gofit %s reached too high "
"baseline error" %
(solver, penalty, gofit))
def __fit_std_Hawkes(dataset_events):
learner = HawkesExpKern([[FITTED_BETA] * NUMBER_OF_DIMENSIONS] *
NUMBER_OF_DIMENSIONS)
learner.fit(dataset_events)
mus = learner.baseline
betas = learner.decays
alphas = learner.adjacency
return mus, alphas, betas
def test_corresponding_simu(self):
"""...Test that the corresponding simulation object is correctly
built
"""
learner = HawkesExpKern(self.decays, max_iter=10)
learner.fit(self.events)
corresponding_simu = learner._corresponding_simu()
self.assertEqual(corresponding_simu.decays, learner.decays)
np.testing.assert_array_equal(corresponding_simu.baseline,
learner.baseline)
np.testing.assert_array_equal(corresponding_simu.adjacency,
learner.adjacency)
def __fit_current_quarter(dataset_events, model_type):
# fit process to current dataset_events
if model_type == "full":
mus, alphas, betas = __fit_std_Hawkes(dataset_events)
elif model_type == "baseline":
mus = np.array([len(dim)
for dim in dataset_events]) / ANALYSIS_PERIOD_OFFSET
alphas = np.zeros((NUMBER_OF_DIMENSIONS, NUMBER_OF_DIMENSIONS))
betas = np.ones(alphas.shape)
elif model_type == "s-e":
mus = []
alphas = []
for dim_i in range(NUMBER_OF_DIMENSIONS):
learner = HawkesExpKern([[FITTED_BETA]])
learner.fit([dataset_events[dim_i]])
mus.append(learner.baseline[0].tolist())
dim_alphas = np.zeros(NUMBER_OF_DIMENSIONS)
dim_alphas[dim_i] = learner.adjacency[0][0].tolist()
alphas.append(dim_alphas)
betas = np.eye(NUMBER_OF_DIMENSIONS) * FITTED_BETA
elif model_type == "reduced":
if EFFECT_TYPE == "s-e":
mus, alphas, betas = __fit_std_Hawkes(dataset_events)
for i in range(NUMBER_OF_DIMENSIONS):
alphas[i][i] = 0
elif EFFECT_TYPE == "late":
mus, alphas, betas = __fit_std_Hawkes(dataset_events)
alphas[1][3] = 0
alphas[3][1] = 0
elif EFFECT_TYPE == "early":
mus, alphas, betas = __fit_std_Hawkes(dataset_events)
alphas[0][2] = 0
alphas[1][2] = 0
alphas[2][0] = 0
alphas[3][0] = 0
alphas[1][0] = 0
else:
raise Exception("unknown model_type")
return mus, alphas, betas
def test_HawkesExpKern_score(self):
"""...Test HawkesExpKern score method
"""
n_nodes = 2
n_realizations = 3
train_events = [[
np.cumsum(np.random.rand(4 + i)) for i in range(n_nodes)
] for _ in range(n_realizations)]
test_events = [[
np.cumsum(np.random.rand(4 + i)) for i in range(n_nodes)
] for _ in range(n_realizations)]
learner = HawkesExpKern(self.decays)
msg = '^You must either call `fit` before `score` or provide events$'
with self.assertRaisesRegex(ValueError, msg):
learner.score()
given_baseline = np.random.rand(n_nodes)
given_adjacency = np.random.rand(n_nodes, n_nodes)
learner.fit(train_events)
train_score_current_coeffs = learner.score()
self.assertAlmostEqual(train_score_current_coeffs, 2.0855840)
train_score_given_coeffs = learner.score(baseline=given_baseline,
adjacency=given_adjacency)
self.assertAlmostEqual(train_score_given_coeffs, 0.59502417)
test_score_current_coeffs = learner.score(test_events)
self.assertAlmostEqual(test_score_current_coeffs, 1.6001762)
test_score_given_coeffs = learner.score(test_events,
baseline=given_baseline,
adjacency=given_adjacency)
self.assertAlmostEqual(test_score_given_coeffs, 0.89322199)
def test_HawkesExpKern_fit_start(self):
"""...Test HawkesExpKern starting point of fit method
"""
n_nodes = len(self.events)
n_coefs = n_nodes + n_nodes * n_nodes
# Do not step
learner = HawkesExpKern(self.decays, max_iter=-1)
learner.fit(self.events)
np.testing.assert_array_equal(learner.coeffs, np.ones(n_coefs))
learner.fit(self.events, start=self.float_1)
np.testing.assert_array_equal(learner.coeffs,
np.ones(n_coefs) * self.float_1)
learner.fit(self.events, start=self.int_1)
np.testing.assert_array_equal(learner.coeffs,
np.ones(n_coefs) * self.int_1)
random_coeffs = np.random.rand(n_coefs)
learner.fit(self.events, start=random_coeffs)
np.testing.assert_array_equal(learner.coeffs, random_coeffs)
def __minimize_loglik_in_beta(event_times_dict_list, betas):
beta_list = [[betas["beta"]] * NUMBER_OF_DIMENSIONS
] * NUMBER_OF_DIMENSIONS
learner = HawkesExpKern(beta_list)
learner.fit(event_times_dict_list)
return learner._solver_obj.get_history()["obj"][-1]
learner = HawkesExpKern(beta_list)
learner.fit(event_times_dict_list)
return learner._solver_obj.get_history()["obj"][-1]
fitted_betas = []
print("starting beta fit")
for iteration_nr in range(BETA_FIT_REPETITIONS):
print(" beta iteration nr" + str(iteration_nr))
fitted_beta = hyperopt.fmin(
fn=lambda betas: __minimize_loglik_in_beta(EVENT_TIMES, betas),
space={"beta": hyperopt.hp.uniform("beta", ZERO, MAXIMUM_BETA)},
algo=hyperopt.tpe.suggest,
max_evals=MAX_HYPEROPT_EVALS)["beta"]
learner = HawkesExpKern([[fitted_beta] * NUMBER_OF_DIMENSIONS] *
NUMBER_OF_DIMENSIONS)
learner.fit(EVENT_TIMES)
fitted_betas.append({
"beta":
fitted_beta,
"loglik":
learner._solver_obj.get_history()["obj"][-1]
})
FITTED_BETA = min(fitted_betas, key=lambda i: i["loglik"])["beta"]
else: # beta fitting can be computationally intensive, so here are the results
FITTED_BETA = 2.288 if MODE == "GROW_VS_DEC" else (
2.067 if MODE == "STEM_VS_HUMAN" else "unknown")
if FITTED_BETA == "unknown":
raise Exception("unknown MODE")
print("beta: {}".format(FITTED_BETA))
for dataset_type in DATASET_LIST.keys():
PERIOD_RESULTS = []
for time_span in range(ANALYSIS_PERIOD_START, ANALYSIS_PERIOD_END + 1):
print("PROCESSING Q{}".format(time_span))
parameter_results = {"mus": [], "alphas": [], "betas": []}
for _ in range(PERMUTED_FITTING_REPETITIONS):
__read_dataset_bound_window = functools.partial(
__read_dataset_window, window_index=time_span)
EVENT_TIMES = POOL.map(__read_dataset_bound_window,
DATASET_LIST[dataset_type])
EVENT_TIMES = [
events for events in EVENT_TIMES if events is not None
]
learner = HawkesExpKern([[FITTED_BETA] * NUMBER_OF_DIMENSIONS] *
NUMBER_OF_DIMENSIONS)
learner.fit(EVENT_TIMES)
parameter_results["mus"].append(
np.array(learner.baseline).tolist())
parameter_results["alphas"].append(
(learner.adjacency * np.array(learner.decays)).tolist())
parameter_results["betas"].append(
np.array(learner.decays).tolist())
PERIOD_RESULTS.append({
"mu": parameter_results["mus"],
"alpha": parameter_results["alphas"],
"beta": parameter_results["betas"],
"#datasets": len(EVENT_TIMES),
"quarter": time_span
})
EVENT_TIMES = None
with open("quarter_permutation_{}_{}.json".format(dataset_type, FIT_TYPE),