def class_by_time(svm_data, grid_2_id):
print 'svm'
trips = []
labels = []
for datas in svm_data:
trip = datas[0]
time = datas[1]
data = []
# 时间特征
for i in time:
data.append(int(i))
p = random.randint(40, 70)
(lat1, lon1, lat_length, lon_length) = gh._decode_c2i(trip[0])
(lat2, lon2, lat_length, lon_length) = gh._decode_c2i(trip[int(len(trip) * float(p) / 100.0)])
data.append(lat1)
data.append(lon1)
data.append(lat2)
data.append(lon2)
print data
trips.append(data)
labels.append(grid_2_id[trip[-1]])
# print len(trips)
# print trips
lin_clf = svm.LinearSVC()
lin_clf.fit(trips, labels)
joblib.dump(lin_clf, "../../data/lin_clf.model")
print 'finish svm'
return lin_clf
def compute_MT(A, M, grid_2_id, id_2_grid):
print 'computing MT'
MT = np.zeros((1584, 1584))
sortlist = defaultdict(list)
for i in id_2_grid:
for j in id_2_grid:
if i == j:
continue
x = id_2_grid[i]
y = id_2_grid[j]
(lat1, lon1, lat_length, lon_length) = gh._decode_c2i(x)
(lat2, lon2, lat_length, lon_length) = gh._decode_c2i(y)
sortlist[(abs(lat1 - lat2) + abs(lon1 - lon2))].append([i, j])
# Mpow = np.eye(1584)
Mpow = gnp.garray(np.eye(1584))
M = gnp.garray(M)
#print Mpow
for i in sortlist:
# print i
Mpow = Mpow.dot(M)
Lde = int(i * 0.2)
Mtemp = Mpow.dot(A[Lde])
# np.dot(Mpow, A[Lde])
# print 'finish'
# print Mtemp
for x in sortlist[i]:
MT[x[0], x[1]] = Mtemp[x[0], x[1]]
# print np.max(MT), np.unravel_index(MT.argmax(), MT.shape)
#print MT
return MT
def subsyn(test_data, trip_data, test_des, M, MT, des, test_time, id_2_grid):
print 'subsyn'
idx = 0
total_error = []
total_km = 0
gtotal_error = []
gtotal_km = 0
for datas in test_data:
# 计算P(Tp)
# Ptp = 100000000
# pre = -1
# for point in datas:
# if pre < 0:
# pre = point
# continue
# Ptp = Ptp * M[pre, point]
#
# 计算P(T^p|d \in n_j)
Ptpnj = defaultdict(float)
for j in des:
if MT[datas[0], j] <= 0:
continue
try:
Ptpnj[j] = MT[datas[-1], j] / MT[datas[0], j]
except:
print ''
P = defaultdict(float)
sum = 0.0
for j in Ptpnj:
P[j] = Ptpnj[j] * des[j]
sum += P[j]
# 改进subsyn
gP = defaultdict(float)
tmp_time = []
for i in test_time[idx]:
tmp_time.append(int(i))
(lat1, lon1, lat_length, lon_length) = gh._decode_c2i(id_2_grid[datas[0]])
(lat2, lon2, lat_length, lon_length) = gh._decode_c2i(id_2_grid[datas[-1]])
predictdata = tmp_time + [lat1, lon1, lat2, lon2]
label = svm_time.predict_by_svm(data=[predictdata])
for j in Ptpnj:
gP[j] = Ptpnj[j] * des[j] * 0.7
if label[0] == j:
gP[j] += 0.3
gP = sorted(gP.iteritems(), key=lambda (k, v): (v, k), reverse=True)
P = sorted(P.iteritems(), key=lambda (k, v): (v, k), reverse=True)
Q = []
for k, v in P:
Q.append(int(k))
gQ = []
for k, v in gP:
gQ.append(int(k))
# print test_des[idx], Q[:5]
# 计算涵盖率
yes_list = []
# print test_des[idx], Q[:5]
if test_des[idx] in Q[:1]:
yes_list.append(1)
else:
yes_list.append(0)
if test_des[idx] in Q[:3]:
yes_list.append(1)
else:
yes_list.append(0)
if test_des[idx] in Q[:5]:
yes_list.append(1)
else:
yes_list.append(0)
# print data, max_ID, max_P, test_des[idx]
total_error.append(yes_list)
# 计算误差曼哈顿距离
# print id_2_grid[int(test_des[idx])]
(lat1, lon1, lat_length, lon_length) = gh._decode_c2i(id_2_grid[int(test_des[idx])])
(lat2, lon2, lat_length, lon_length) = gh._decode_c2i(id_2_grid[Q[0]])
total_km += abs(lat1 - lat2) + abs(lon1 - lon2)
# 计算改进subsyn
# print test_des[idx], Q[:5]
# 计算涵盖率
yes_list = []
# print test_des[idx], Q[:5]
if test_des[idx] in gQ[:1]:
yes_list.append(1)
else:
yes_list.append(0)
if test_des[idx] in gQ[:3]:
yes_list.append(1)
else:
yes_list.append(0)
if test_des[idx] in gQ[:5]:
yes_list.append(1)
else:
yes_list.append(0)
# print data, max_ID, max_P, test_des[idx]
gtotal_error.append(yes_list)
# 计算误差曼哈顿距离
# print id_2_grid[int(test_des[idx])]
(lat1, lon1, lat_length, lon_length) = gh._decode_c2i(id_2_grid[int(test_des[idx])])
(lat2, lon2, lat_length, lon_length) = gh._decode_c2i(id_2_grid[gQ[0]])
gtotal_km += abs(lat1 - lat2) + abs(lon1 - lon2)
idx += 1
P1 = P3 = P5 = 0.0
for data in total_error:
P1 += data[0]
P3 += data[1]
P5 += data[2]
print P1, P3, P5
print P1 / len(test_des), P3 / len(test_des), P5 / len(test_des)
print total_km * 1.0 / len(test_des)
P1 = P3 = P5 = 0.0
for data in gtotal_error:
P1 += data[0]
P3 += data[1]
P5 += data[2]
print P1, P3, P5
print P1 / len(test_des), P3 / len(test_des), P5 / len(test_des)
print gtotal_km * 1.0 / len(test_des)
def ZMDB(test_data, trip_data, test_des, id_2_grid):
print 'ZMDB'
# print test_des
total_error = []
total_km = 0
try:
idx = 0
for data in test_data:
des_num = defaultdict(int)
num = 0
# 计算 每个查询轨迹:满足目的地为n^j 且 查询轨迹匹配trip_data的数目:des_num 总的匹配轨迹为num
for trip in trip_data:
tmp = testINtrip(data, trip)
num += tmp
des_num[trip[-1]] += tmp
if num <= 0:
idx += 1
continue
# print num
max_P = -1
max_ID = -1
# 对每个目的地而言, P(n^j | T^end(np.eye(1584))p)
P = defaultdict(float)
for i in des_num:
P[i] = des_num[i] * 1.0 / num
if max_P < P[i]:
max_P = P[i]
max_ID = i
P = sorted(P.iteritems(), key=lambda (k, v): (v, k), reverse=True)
Q = []
for k, v in P:
Q.append(int(k))
# 计算涵盖率
yes_list = []
# print test_des[idx], Q[:5]
if test_des[idx] in Q[:1]:
yes_list.append(1)
else:
yes_list.append(0)
if test_des[idx] in Q[:3]:
yes_list.append(1)
else:
yes_list.append(0)
if test_des[idx] in Q[:5]:
yes_list.append(1)
else:
yes_list.append(0)
# print data, max_ID, max_P, test_des[idx]
total_error.append(yes_list)
# 计算误差曼哈顿距离
# print test_des[idx]
# print id_2_grid[int(test_des[idx])]
(lat1, lon1, lat_length, lon_length) = gh._decode_c2i(id_2_grid[int(test_des[idx])])
(lat2, lon2, lat_length, lon_length) = gh._decode_c2i(id_2_grid[Q[0]])
total_km += abs(lat1 - lat2) + abs(lon1 - lon2)
idx += 1
except:
# print test_des[idx], Q[0]
s = sys.exc_info()
print "Error '%s' happened on line %d" % (s[1], s[2].tb_lineno)
P1 = P3 = P5 = 0.0
for data in total_error:
P1 += data[0]
P3 += data[1]
P5 += data[2]
print P1, P3, P5
print P1 / len(test_des), P3 / len(test_des), P5 / len(test_des)
print total_km * 1.0 / len(test_des)