def train_strat_model(weights, data_train, data_val, data_test, lambd):
loss = strat_models.nonparametric_discrete_loss()
reg = strat_models.scaled_plus_sum_squares_reg(A=D, lambd=lambd)
bm = strat_models.BaseModel(loss=loss, reg=reg)
G_week = nx.cycle_graph(53)
G_hr = nx.cycle_graph(24)
strat_models.set_edge_weight(G_week, weights[0])
strat_models.set_edge_weight(G_hr, weights[1])
G = strat_models.cartesian_product([G_week, G_hr])
sm = strat_models.StratifiedModel(bm, graph=G)
info = sm.fit(data_train, **kwargs)
anll_train = sm.anll(data_train)
anll_val = sm.anll(data_val)
anll_test = sm.anll(data_test)
print("Stratified model with (weights, lambd) =", (weights, lambd))
print("\t", info)
print("\t", anll_train, anll_val, anll_test)
return anll_train, anll_val, anll_test
def make_G(w1, w2, w3):
G_vix = nx.path_graph(10) #vix quantiles (deciles)
G_unemp = nx.path_graph(10) #volume quantiles
G_inflation = nx.path_graph(10) #volume quantiles
strat_models.set_edge_weight(G_vix, w1)
strat_models.set_edge_weight(G_unemp, w2)
strat_models.set_edge_weight(G_inflation, w3)
G = strat_models.cartesian_product([G_vix, G_unemp, G_inflation])
return G.copy()
for state2 in states:
if state2 in list(
neighbors[neighbors.StateCode == state1]['NeighborStateCode']):
G_state.add_edge(state1, state2)
n_years = len(years)
G_time = nx.path_graph(n_years)
G_time = nx.relabel_nodes(G_time, dict(zip(np.arange(n_years), years)))
kwargs = dict(abs_tol=1e-5, rel_tol=1e-5, maxiter=200, n_jobs=4, verbose=1)
loss = strat_models.bernoulli_loss()
reg = strat_models.clip_reg(lambd=(1e-5, 1 - 1e-5))
bm = strat_models.BaseModel(loss=loss, reg=reg)
strat_models.set_edge_weight(G_state, 0)
strat_models.set_edge_weight(G_time, 0)
G = strat_models.cartesian_product([G_state, G_time])
sm_fully = strat_models.StratifiedModel(bm, graph=G)
info = sm_fully.fit(data_train, **kwargs)
anll_train = sm_fully.anll(data_train)
anll_test = sm_fully.anll(data_test)
print("Separate model")
print("\t", info)
print("\t", anll_train, anll_test)
strat_models.set_edge_weight(G_state, 1)
strat_models.set_edge_weight(G_time, 4)
G = strat_models.cartesian_product([G_state, G_time])
sm_strat = strat_models.StratifiedModel(bm, graph=G)
data_train = dict(X=X_train, Y=Y_train, Z=Z_train)
data_test = dict(X=X_test, Y=Y_test, Z=Z_test)
loss = strat_models.logistic_loss(intercept=True)
# Fit models
print("fitting...")
kwargs = dict(rel_tol=1e-4,
abs_tol=1e-4,
maxiter=500,
n_jobs=12,
verbose=True,
rho=2.,
max_cg_iterations=30)
strat_models.set_edge_weight(G_sex, 0)
strat_models.set_edge_weight(G_age, 0)
G = strat_models.utils.cartesian_product([G_sex, G_age])
bm_fully = strat_models.BaseModel(loss=loss)
sm_fully = strat_models.StratifiedModel(bm_fully, graph=G)
info = sm_fully.fit(data_train, **kwargs)
anll_test = sm_fully.anll(data_test)
pred_error = prediction_error(data_test, sm_fully)
print('Separate model')
print('\t', info)
print('\t', anll_test, pred_error)
strat_models.set_edge_weight(G_sex, 10)
Z = []
for node in G.nodes():
Y.append(events[node])
Z.append(node)
return Y, Z
Y_train, Z_train = df_to_data(df_2017)
Y_test, Z_test = df_to_data(df_2018)
print(len(Y_train), len(Y_test))
del G
# Fit models and evaluate log likelihood
kwargs = dict(rel_tol=1e-6, abs_tol=1e-6, maxiter=2000)
strat_models.set_edge_weight(G_location, 0)
strat_models.set_edge_weight(G_week, 0)
strat_models.set_edge_weight(G_day, 0)
strat_models.set_edge_weight(G_hour, 0)
G = strat_models.cartesian_product([G_location, G_week, G_day, G_hour])
fully = strat_models.Poisson()
info = fully.fit(Y_train, Z_train, G, **kwargs)
anll_train = fully.anll(Y_train, Z_train)
anll_test = fully.anll(Y_test, Z_test)
print("Fully")
print("\t", info)
print("\t", anll_train, anll_test)
del G
strat_models.set_edge_weight(G_location, 100)
strat_models.set_edge_weight(G_week, 100)
# Fit models
data_train = dict(X=X_train, Y=Y_train, Z=Z_train)
data_test = dict(X=X_test, Y=Y_test, Z=Z_test)
kwargs = dict(rel_tol=1e-5, abs_tol=1e-5, maxiter=1000, n_jobs=2, verbose=1)
def rms(x):
return np.sqrt(np.mean(np.square(x)))
loss = strat_models.sum_squares_loss(intercept=True)
reg = strat_models.sum_squares_reg(lambd=1e-4)
bm = strat_models.BaseModel(loss=loss, reg=reg)
strat_models.set_edge_weight(G, 1e-8)
sm_fully = strat_models.StratifiedModel(bm, graph=G)
info = sm_fully.fit(data_train, **kwargs)
score = sm_fully.scores(data_test)
print("Fully")
print("\t", info)
print("\t", score)
strat_models.set_edge_weight(G, 15)
sm_strat = strat_models.StratifiedModel(bm, graph=G)
info = sm_strat.fit(data_train, **kwargs)
score = sm_strat.scores(data_test)
print("Strat")
## Eigen-stratified model
print("fitting eigen-stratified models...")
kwargs["maxiter"] = 600
kwargs["verbose"] = False
K = 53 * 24
weight_week = .45
weight_hr = .55
lambd = (0.01, 0.001)
m = 90
G_week = nx.cycle_graph(53)
G_hr = nx.cycle_graph(24)
strat_models.set_edge_weight(G_week, weight_week)
strat_models.set_edge_weight(G_hr, weight_hr)
G_eigen = strat_models.cartesian_product([G_week, G_hr])
loss = strat_models.nonparametric_discrete_loss()
reg = strat_models.scaled_plus_sum_squares_reg(A=D, lambd=lambd)
bm_eigen = strat_models.BaseModel(loss=loss, reg=reg)
sm_eigen = strat_models.StratifiedModel(bm_eigen, graph=G_eigen)
info = sm_eigen.fit(data_train, num_eigen=m, **kwargs)
anll_train = sm_eigen.anll(data_train)
anll_val = sm_eigen.anll(data_val)
anll_test = sm_eigen.anll(data_test)
print('Eigen-stratified model, {} eigenvectors used'.format(m))