def var_sepe(data_):
ds = data_.shape
data, miss_data1, W1, ori_W1 = handle_data(data_)
lou, K, conv_thre, fb = 3e-3, 100, 1e-5, 0.85
est_ori = halrtc_cpt(miss_data1, lou, conv_thre, K, W1, fb)
var_list = []
mean_list, svd_list = [], []
for r in range(ds[0]):
var_list.append(np.var(data[r]))
mean_list.append(np.mean(data[r]))
var_dict = {var_list[i]: i for i in range(ds[0])}
var_assign = [[], [], []]
for var in var_dict:
if var < 50:
var_assign[0].append(var_dict[var])
elif var < 150:
var_assign[1].append(var_dict[var])
else:
var_assign[2].append(var_dict[var])
est_var = np.zeros_like(data)
for va in var_assign:
est_va = halrtc_cpt(miss_data1[va], lou, conv_thre, K, W1[va], fb)
est_var[va] = est_va
eva_var = rmse_mape_rse(est_var, data, (W1 | (ori_W1 == False)))
print('sepe', eva_var)
est_ori = halrtc_cpt(miss_data1, lou, conv_thre, K, W1, fb)
eva_ori = rmse_mape_rse(est_ori, data, (W1 | (ori_W1 == False)))
print('ori', eva_ori)
return
def week_sepe(handle_train, handle_test):
data, miss_data1, W1, ori_W1 = handle_train
halrtc_para = [lou, K, conv_thre, fb] = [3e-3, 100, 1e-5]
data_test, miss_data_test, Wtest, ori_Wtest = handle_test
ds_test = data_test.shape
est_ori = halrtc_cpt(miss_data_test, lou, conv_thre, K, Wtest, fb)
weekday2, weekend2 = [], []
for i in range(ds_test[1]):
if i % 7 < 5:
weekday2.append(i)
else:
weekend2.append(i)
est1 = halrtc_cpt(miss_data_test[:, weekday2, :], lou, conv_thre, K,
Wtest[:, weekday2, :], fb)
eva1 = rmse_mape_rse(est1, data_test[:, weekday2, :],
(Wtest | (ori_Wtest == False))[:, weekday2, :])
est2 = halrtc_cpt(miss_data_test[:, weekend2, :], lou, conv_thre, K,
Wtest[:, weekend2, :], fb)
eva2 = rmse_mape_rse(est2, data_test[:, weekend2, :],
(Wtest | (ori_Wtest == False))[:, weekend2, :])
print('weekday', eva1)
print(
'ori_weekday',
rmse_mape_rse(est_ori[:, weekday2, :], data_test[:, weekday2, :],
(Wtest | (ori_Wtest == False))[:, weekday2, :]))
print('weekend', eva2)
print(
'ori_weekend',
rmse_mape_rse(est_ori[:, weekend2, :], data_test[:, weekend2, :],
(Wtest | (ori_Wtest == False))[:, weekend2, :]))
return
def compare_3d_4d(handle_train, handle_test):
data, miss_data1, W1, ori_W1 = handle_train
halrtc_para = [lou, K, conv_thre, fb] = [1e-3, 100, 1e-5]
data_test, miss_data_test, Wtest, ori_Wtest = handle_test
rW = Wtest | (ori_Wtest == False)
data_size = data_test.shape
weeks = data_size[1] // 7
'''
Nori_data = np.zeros((data_size[0], weeks, 7, data_size[2]))
Nmiss_data = np.zeros_like(Nori_data)
N_W = np.zeros_like(Nmiss_data)
Nori_W = np.zeros_like(Nori_data)
Nr_W = np.zeros_like(Nori_W)
for i in range(data_size[1]):
if i >= weeks * 7:
break
Nori_data[:, i // 7, i % 7, :] = data_test[:, i, :]
Nori_W[:, i // 7, i % 7, :] = ori_Wtest[:, i, :]
Nmiss_data[:, i // 7, i % 7, :] = miss_data_test[:, i, :]
N_W[:, i // 7, i % 7, :] = Wtest[:, i, :]
Nr_W[:, i // 7, i % 7, :] = rW[:, i, :]
print(Nmiss_data.shape)
time0 = time.time()
est_4d = np.zeros_like(Nmiss_data)
est_4d[L==0] = halrtc_cpt(Nmiss_data[L==0], 1e-3, 1e-4, 100, N_W[L==0], 0)
est_4d[L == 1] = halrtc_cpt(Nmiss_data[L == 1], 1e-3, 1e-4, 100, N_W[L==1], 0)
time1 = time.time()
print('4d_halrtc:', rmse_mape_rse(est_4d, Nori_data, Nr_W))
print('4d_time', str(time1 - time0) + 's')
'''
print('loss', (miss_data_test < 1).sum())
time0 = time.time()
est_ori = halrtc_cpt(miss_data_test, lou, conv_thre, K, Wtest, fb)
time1 = time.time()
eva_ori = rmse_mape_rse(est_ori, data_test, rW)
print('ori:', eva_ori)
time_ori = time1 - time0
print('ori_time:', str(time_ori) + 's')
return
L = np.zeros(data_size[0])
var_mat, mean_mat = traffic_info(handle_train)
for i in range(data_size[0]):
#按方差均值划分
t = np.var(data[i])
if t < 50:
s = round(t**0.5 / np.mean(data[i]), 2)
L[i] = 1
#L = var_cluster(handle_train)
time0 = time.time()
est_3d = cluster_ha(L, miss_data_test, Wtest, 2, halrtc_para)
time1 = time.time()
eva_3d = rmse_mape_rse(est_3d, data_test, rW)
print('3d:', eva_3d)
time_3d = time1 - time0
print('3d_time:', str(time_3d) + 's')
return eva_ori, time_ori, eva_3d, time_3d
def test_simM(data_, r, dis_matrix, k):
r_dis = dis_matrix[r]
dis_dict = {r_dis[i]: i for i in range(len(r_dis))}
sim_road = [dis_dict[j] for j in sorted(dis_dict.keys())[:k + 1]]
data, miss_data1, W1, ori_W1 = handle_data(data_)
lou, K, conv_thre, fb = 3e-3, 100, 1e-5, 0.85
est_ori = halrtc_cpt(miss_data1, lou, conv_thre, K, W1, fb)
eva_ori = rmse_mape_rse(est_ori[r], data[r], (W1 | (ori_W1 == False))[r])
est_sim = halrtc_cpt(miss_data1[sim_road], lou, conv_thre, K, W1[sim_road],
fb)
eva_sim = rmse_mape_rse(est_sim[0], data[r], (W1 | (ori_W1 == False))[r])
print(eva_ori)
print(eva_sim)
return
def test_labels(cn, labels, handle_info):
halrtc_para = [lou, K, conv_thre, fb] = [3e-3, 100, 1e-5]
data_test, miss_data_test, Wtest, ori_Wtest = handle_info
ds_test = data_test.shape
est_ori = halrtc_cpt(miss_data_test, lou, conv_thre, K, Wtest, fb)
for i in range(cn):
print('label-' + str(i), (labels == i).sum())
temp_data = halrtc_cpt(miss_data_test[labels == i], lou, conv_thre, K,
Wtest[labels == i], fb)
#est_c = cluster_ha(labels, miss_data_test, Wtest, cn, halrtc_para, axis=0)
eva_c = rmse_mape_rse(temp_data, data_test[labels == i],
(Wtest | (ori_Wtest == False))[labels == i])
print('clu', eva_c)
eva_test = rmse_mape_rse(est_ori[labels == i], data_test[labels == i],
(Wtest | (ori_Wtest == False))[labels == i])
print('ori', eva_test)
return
def test_week_cluster(handle_train, handle_test, choise='var'):
#按一周七天分别聚类
data, miss_data1, W1, ori_W1 = handle_train
ds = data.shape
train = {}
halrtc_para = [lou, K, conv_thre, fb] = [3e-3, 100, 1e-5]
data_test, miss_data_test, Wtest, ori_Wtest = handle_test
est_ori = halrtc_cpt(miss_data_test, lou, conv_thre, K, Wtest, fb)
ds_test = data_test.shape
test, est_clu = {}, {}
est_week = np.zeros_like(data_test)
K_n, max_d = 2, 100
for i in range(7):
train[i], test[i] = [], []
days_train = []
for k in range(ds[1]):
if k % 7 == i:
days_train.append(k)
for s in handle_train:
train[i].append(s[:, days_train, :])
days_test = []
for k in range(ds_test[1]):
if k % 7 == i:
days_test.append(k)
for t in handle_test:
test[i].append(t[:, days_test, :])
L = np.zeros(ds[0])
for j in range(ds[0]):
if np.var(data[j, i]) > 50:
L[j] = 1
#L = SC_1(train[i][0], 6, K_n, method='pearson')
est_clu[i] = cluster_ha(L, test[i][1], test[i][2], K_n, halrtc_para)
#est_clu[i] = test_cluster(train[i], test[i])[-1]
for p in range(est_clu[i].shape[1]):
est_week[:, p * 7 + i, :] = est_clu[i][:, p, :]
eva_week = rmse_mape_rse(est_clu[i], data_test[:, days_test, :],
(Wtest | (ori_Wtest == False))[:,
days_test, :])
print('clu', i, eva_week)
eva_i = rmse_mape_rse(est_ori[:, days_test, :],
data_test[:, days_test, :],
(Wtest | (ori_Wtest == False))[:, days_test, :])
print('ori', i, eva_i)
return
def clusters_train(handle_train, handle_test):
K_n = 2
halrtc_para = [lou, K, conv_thre, fb] = [1e-3, 100, 1e-5]
data, miss_data1, W1, ori_W1 = handle_train
data_test, miss_data_test, Wtest, ori_Wtest = handle_test
est_ori = halrtc_cpt(miss_data_test, lou, conv_thre, K, Wtest, fb)
time0 = time.time()
var_mat, mean_mat = traffic_info(handle_test)
clf = KMeans(n_clusters=K_n)
S = clf.fit(var_mat)
L = S.labels_
print(L)
'''
L = np.zeros(209)
var_list = []
std_list = []
var_mean = np.zeros((209,2))
for i in range(209):
#按方差均值划分
#if np.mean(var_mat[i]) > 50:
#t = np.std(data[i])/np.mean(data[i])
t = np.mean(var_mat[i])
var_list.append(t)
var_mean[0] = np.var(data[i])
var_mean[1] = np.mean(data[i])
if t < 50:
s = round(t**0.5/np.mean(data[i]),2)
L[i] = 1
return L
'''
#est_Kmeans[L==1] = est_ori[L==1]
est_Kmeans = cluster_ha(L, miss_data_test, Wtest, K_n, halrtc_para)
time1 = time.time()
for i in range(K_n):
print(
'ori_' + str(i),
rmse_mape_rse(est_ori[L == i], data_test[L == i],
(Wtest | (ori_Wtest == False))[L == i]))
print(
'KM_' + str(i),
rmse_mape_rse(est_Kmeans[L == i], data_test[L == i],
(Wtest | (ori_Wtest == False))[L == i]))
print('Kmeans:',
rmse_mape_rse(est_Kmeans, data_test, Wtest | (ori_Wtest == False)))
print('Kmeans_time:', str(time1 - time0) + 's')
time0 = time.time()
est_ori = halrtc_cpt(miss_data_test, lou, conv_thre, K, Wtest, fb)
time1 = time.time()
eva_ori = rmse_mape_rse(est_ori, data_test, (Wtest | (ori_Wtest == False)))
print('ori:', eva_ori)
print('ori_time:', str(time1 - time0) + 's')
'''
time0 = time.time()
L = SC_1(data,10,K_n,method='pearson')
est_SC = cluster_ha(L,miss_data_test,Wtest,K_n,halrtc_para)
time1 = time.time()
print('SC:',rmse_mape_rse(est_SC,data_test,Wtest|(ori_Wtest==False)))
print('SC_time:',str(time1-time0)+'s')
time0 = time.time()
L = road_Kmeans(data,ori_W1,K_n,W1,method='pearson')[0]
est_Kmeans = cluster_ha(L, miss_data_test, Wtest, K_n,halrtc_para)
time1 = time.time()
print('Kmeans:', rmse_mape_rse(est_Kmeans, data_test, Wtest | (ori_Wtest == False)))
print('Kmeans_time:',str(time1-time0)+'s')
'''
return
def test_cluster(handle_train, handle_test, choise='var'):
p_tr = 0.9
halrtc_para = [lou, K, conv_thre, fb] = [1.3e-3, 100, 1e-5]
data_test, miss_data_test, Wtest, ori_Wtest = handle_test
rW = Wtest | (ori_Wtest == False)
time_a = time.time()
#lou = 1/(3*T_SVD(miss_data_test,p_tr)[0])
est_ori = halrtc_cpt(miss_data_test, lou, conv_thre, K, Wtest, fb)
eva_ori = rmse_mape_rse(est_ori, data_test, (Wtest | (ori_Wtest == False)))
K_n = 2
max_d = 100
time_s = time.time()
ori_time = time_s - time_a
print('ori', eva_ori)
print('ori_time', str(ori_time) + 's')
ts = data_test.shape
L = np.zeros(ts[0])
Road = np.array(range(ts[0]))
max_c = 0
est_clu = np.zeros_like(data_test)
#handle_train = handle_test #change
var_mat, mean_mat = traffic_info(handle_train)
iter = 0
while True:
HD_test = []
for temp in handle_test:
HD_test.append(temp[L == max_c])
var_mat = var_mat[L == max_c]
mean_mat = mean_mat[L == max_c]
L, c = AHC(handle_train, choise, var_mat, mean_mat, K_n, max_d)
if c < 0.65 or iter > 3:
break
#test_labels(K_n,L,HD_test)
HD_train = []
max_c, max_n = -1, 0
for i in range(K_n):
if (L == i).sum() > max_n:
max_c = i
max_n = (L == i).sum()
for i in range(K_n):
if i != max_c:
list_i = Road[L == i]
lou = 1 / (3 * T_SVD(miss_data_test[list_i], p_tr)[0])
est_clu[list_i] = halrtc_cpt(miss_data_test[list_i], lou,
conv_thre, K, Wtest[list_i], fb)
print(
'clu',
rmse_mape_rse(est_clu[list_i], data_test[list_i],
rW[list_i]))
print(
'ori',
rmse_mape_rse(est_ori[list_i], data_test[list_i],
rW[list_i]))
for temp in handle_train:
HD_train.append(temp[L == max_c])
handle_train = HD_train.copy()
handle_test = HD_test.copy()
Road = Road[L == max_c]
iter += 1
#print(Road)
#if Road.size < 80:
# break
#最后一部分聚类与否
clf = KMeans(n_clusters=K_n)
data, miss_data1, W1, ori_W1 = HD_train
ds1 = data.shape
mean_mat = np.zeros(ds1[:2])
for r in range(ds1[0]):
for d in range(ds1[1]):
mean_mat[r, d] = np.mean(data[r, d, :])
#mean_mat[r, d] = np.std(data[r,d,:])/np.mean(data[r,d,:])
s = clf.fit(mean_mat)
L = np.array(s.labels_)
for i in range(K_n):
list_i = Road[L == i]
#lou = 1 / (3 * T_SVD(miss_data_test[list_i], p_tr)[0])
est_clu[list_i] = halrtc_cpt(miss_data_test[list_i], lou, conv_thre, K,
Wtest[list_i], fb)
#lou = 1 / (3 * T_SVD(miss_data_test[Road], p_tr)[0])
#est_clu[Road] = halrtc_cpt(miss_data_test[Road], lou, conv_thre, K, Wtest[Road], fb)
eva_clu = rmse_mape_rse(est_clu, data_test, (Wtest | (ori_Wtest == False)))
time_e = time.time()
ori_time = time_s - time_a
print('ori', eva_ori)
print('ori_time', str(ori_time) + 's')
print('clu', eva_clu)
clu_time = time_e - time_s
print('clu_time', str(clu_time) + 's')
return eva_ori, ori_time, eva_clu, clu_time, est_clu