def test_mean_squared_error(self):
try:
from sklearn.metrics import mean_squared_error as skmse
except:
unittest.TestCase.skipTest(self, "sklearn is not found in the libraries")
skmse_score1 = skmse(self.local_reg1.target, self.local_reg1.p_target)
dlpymse_score1 = mean_squared_error('target', 'p_target', castable=self.reg_table1)
self.assertAlmostEqual(skmse_score1, dlpymse_score1)
skmse_score2 = skmse(self.local_reg1.target, self.local_reg2.p_target)
dlpymse_score2 = mean_squared_error(self.reg_table1.target,self.reg_table2.p_target,
id_vars='id1')
self.assertAlmostEqual(skmse_score2, dlpymse_score2)
def test_mean_squared_error(self):
try:
from sklearn.metrics import mean_squared_error as skmse
except:
unittest.TestCase.skipTest(self, "sklearn is not found in the libraries")
local_reg1 = self.reg_table1.to_frame()
skmse_score1 = skmse(local_reg1.target, local_reg1.p_target)
dlpymse_score1 = mean_squared_error(self.reg_table1, 'target', 'p_target')
self.assertAlmostEqual(skmse_score1, dlpymse_score1)
K_VALS = [3, 5, 7, 9, 11, 13, 15]
starttime = time.time()
# Repeat each trial 10 times.
for i in range (0, 10):
x_train, x_test, y_train, y_test = train_test_split(X, y,\
test_size=0.2)
"""
Try non-optimized methods.
"""
# Vanilla KNN.
for k in K_VALS:
reg = KNNRegressor(x_train, y_train, k)
y_pred = reg.predict(x_test)
mse_iter = skmse(y_test, y_pred)
print("xx,knn,", k,",", mse_iter)
# Distance-weighted KNN.
for k in K_VALS:
reg = DwKNNRegressor(x_train, y_train, k)
y_pred = reg.predict(x_test)
mse_iter = skmse(y_test, y_pred)
print("xx,dknn,", k,",", mse_iter)
"""
PCA with KNN.
"""
pca = PCA(n_components = 6)
pca.fit(x_train.copy())
x_train_pca = pca.transform(x_train.copy())
weights[X_ids] += reweights * A
# Results with optimized subsampling
best_idx = np.random.choice(len(x_train_p),\
size = int(p * n_samples),\
replace = False,\
p = weights/weights.sum())
X_train_p = x_train_p[best_idx]
y_train_p = y_train[best_idx]
clf = neighbors.KNeighborsRegressor(k)
clf.fit(X_train_p[:, :-1], y_train_p)
y_pred = clf.predict(x_test)
print("p-sampling,", k, ",", p, ",", skmse(y_pred, y_test))
# Uniform subsampling.
best_idx = np.random.choice(len(x_train_p),\
size = int(p * n_samples),\
replace = False)
X_train_p = x_train_p[best_idx]
y_train_p = y_train[best_idx]
clf = neighbors.KNeighborsRegressor(k)
clf.fit(X_train_p[:, :-1], y_train_p)
y_pred = clf.predict(x_test)
print("u-sampling,", ",", k, ",", p, ",", skmse(y_pred, y_test))
def ga_run(x_train, y_train, x_test, y_test, x_verif, y_verif, k):
# Run GA to find best weights.
N_init_pop = 50
N_crossover = 50
N_selection = 20
improv_thresh = 1e-3
_, nFeats = np.shape(x_train)
weight_ga = GeneticAlgorithm(nFeats, N_init_pop, mu=0.1)
weight_pop = weight_ga.get_population()
metric_array = np.empty(N_init_pop)
# Create the initial population.
for i in range(len(weight_pop)):
# Scale input data
scaled_x_train = np.multiply(x_train, weight_pop[i])
# Scale verificaion data
scaled_x_verif = np.multiply(x_verif, weight_pop[i])
# Regressor.
reg = KNNRegressor(scaled_x_train, y_train, k)
neighbors = reg.find_all_neighbors(scaled_x_verif)
nbh_std = reg.find_neighborhood_std(neighbors)
metric_array[i] = nbh_std
# Update fitness in GA object.
weight_ga.set_fitness(metric_array)
weight_ga.selection(N_selection)
new_best_metric = 2.5
# while (best_metric - new_best_metric) > improv_thresh:
count = 0
while (count < 20):
count += 1
best_metric = new_best_metric
# Crossover.
weight_ga.crossover(N_crossover)
# Get new population.
weight_pop = weight_ga.get_population()
metric_array = np.empty(N_crossover)
# Evaluate and set fitness.
for i in range(len(weight_pop)):
# Scale input data
scaled_x_train = np.multiply(x_train, weight_pop[i])
# Scale verificaion data
scaled_x_verif = np.multiply(x_verif, weight_pop[i])
# Regressor.
reg = KNNRegressor(scaled_x_train, y_train, k)
neighbors = reg.find_all_neighbors(scaled_x_verif)
nbh_std = reg.find_neighborhood_std(neighbors)
metric_array[i] = nbh_std
# Update fitness in GA object
weight_ga.set_fitness(metric_array)
# get_best_sol
best_weights, new_best_metric = weight_ga.best_sol()
#print("Metric of this iteration are: ", new_best_metric)
weight_ga.selection(N_selection)
# print("Best weights = ", best_weights, "\tBest metric = ", new_best_metric)
# Test with scaling after GA
# Concatenate training and verification sets.
x_train = np.concatenate((x_train, x_verif), axis=0)
y_train = np.concatenate([y_train, y_verif])
# Print the results of KNN.
reg = KNNRegressor(np.multiply(x_train, best_weights), y_train, k)
y_pred = reg.predict(np.multiply(x_test, best_weights))
mse_iter = skmse(y_test, y_pred)
print("ga,knn,", k, ",", mse_iter)
# Print the results of KNN.
reg = DwKNNRegressor(np.multiply(x_train, best_weights), y_train, k)
y_pred = reg.predict(np.multiply(x_test, best_weights))
mse_iter = skmse(y_test, y_pred)
print("ga,dknn,", k, ",", mse_iter)
def lbest_pso_run(x_train, y_train, x_test, y_test, x_verif, y_verif, k):
#Run PSO to find best weights
N_init_pop = 50
_, nFeats = np.shape(x_train)
weight_pso = LBestPSO(nFeats, N_init_pop)
pos = weight_pso.get_positions()
pbest = weight_pso.get_pbest()
pbest_metric_array = np.empty(N_init_pop)
pos_metric_array = np.empty(N_init_pop)
#Set pbest metrics
for i in range(len(pbest)):
#Scale input data
scaled_x_train = np.multiply(x_train, pbest[i])
#Scale verificaion data
scaled_x_verif = np.multiply(x_verif, pbest[i])
#Method 1
reg = KNNRegressor(scaled_x_train, y_train, k)
neighbors = reg.find_all_neighbors(scaled_x_verif)
nbh_std = reg.find_neighborhood_std(neighbors)
pbest_metric_array[i] = nbh_std
weight_pso.set_pbest_fitness(pbest_metric_array)
#Set pos metrics
for i in range(len(pbest)):
#Scale input data
scaled_x_train = np.multiply(x_train, pos[i])
#Scale verificaion data
scaled_x_verif = np.multiply(x_verif, pos[i])
#Method 1
reg = KNNRegressor(scaled_x_train, y_train, k)
neighbors = reg.find_all_neighbors(scaled_x_verif)
nbh_std = reg.find_neighborhood_std(neighbors)
pos_metric_array[i] = nbh_std
weight_pso.set_p_fitness(pos_metric_array)
#Set initial gbest.
weight_pso.set_init_best(pos_metric_array)
count = 0
while (count < 50):
count += 1
weight_pso.optimize()
#get_population
weight_pop = weight_pso.get_positions()
metric_array = np.empty(N_init_pop)
#evaluate and set fitness
for i in range(len(weight_pop)):
#Scale input data
scaled_x_train = np.multiply(x_train, weight_pop[i])
#Scale verificaion data
scaled_x_verif = np.multiply(x_verif, weight_pop[i])
#Method 1
reg = KNNRegressor(scaled_x_train, y_train, k)
neighbors = reg.find_all_neighbors(scaled_x_verif)
nbh_std = reg.find_neighborhood_std(neighbors)
metric_array[i] = nbh_std
weight_pso.set_p_fitness(metric_array)
weight_pso.set_best(metric_array)
#get_best_sol
best_metric = weight_pso.get_gbest_fit()
best_weights = weight_pso.get_gbest()
# Concatenate training and verification sets.
x_train = np.concatenate((x_train, x_verif), axis=0)
y_train = np.concatenate([y_train, y_verif])
# Print the results of KNN.
reg = KNNRegressor(np.multiply(x_train, best_weights), y_train, k)
y_pred = reg.predict(np.multiply(x_test, best_weights))
mse_iter = skmse(y_test, y_pred)
print("lbest-pso,knn,", k, ",", mse_iter)
# Print the results of KNN.
reg = DwKNNRegressor(np.multiply(x_train, best_weights), y_train, k)
y_pred = reg.predict(np.multiply(x_test, best_weights))
mse_iter = skmse(y_test, y_pred)
print("lbest-pso,dknn,", k, ",", mse_iter)