"""
数据预处理阶段相关的工具函数
"""
import numpy as np
import time
from datetime import datetime, timedelta
from collections import defaultdict
[docs]def parse_time(time_in, timezone_offset_in_minute=0):
"""
将 json 中 time_format 格式的 time 转化为 local datatime
"""
date = datetime.strptime(time_in, '%Y-%m-%dT%H:%M:%SZ') # 这是 UTC 时间
return date + timedelta(minutes=timezone_offset_in_minute)
[docs]def cal_basetime(start_time, base_zero):
"""
用于切分轨迹成一个 session，
思路为：给定一个 start_time 找到一个基准时间 base_time，
在该 base_time 到 base_time + time_length 区间的点划分到一个 session 内，
选取 base_time 来做的理由是：这样可以保证同一个小时段总是被 encode 成同一个数
"""
if base_zero:
return start_time - timedelta(hours=start_time.hour,
minutes=start_time.minute,
seconds=start_time.second,
microseconds=start_time.microsecond)
else:
# time length = 12
if start_time.hour < 12:
return start_time - timedelta(hours=start_time.hour,
minutes=start_time.minute,
seconds=start_time.second,
microseconds=start_time.microsecond)
else:
return start_time - timedelta(hours=start_time.hour - 12,
minutes=start_time.minute,
seconds=start_time.second,
microseconds=start_time.microsecond)
[docs]def cal_timeoff(now_time, base_time):
"""
计算两个时间之间的差值，返回值以小时为单位
"""
# 先将 now 按小时对齐
delta = now_time - base_time
return delta.days * 24 + delta.seconds / 3600
[docs]def caculate_time_sim(data):
time_checkin_set = defaultdict(set)
tim_size = data['tim_size']
data_neural = data['data']
for uid in data_neural:
uid_sessions = data_neural[uid]
for session in uid_sessions:
for checkin in session:
timid = checkin[1]
locid = checkin[0]
if timid not in time_checkin_set:
time_checkin_set[timid] = set()
time_checkin_set[timid].add(locid)
sim_matrix = np.zeros((tim_size, tim_size))
for i in range(tim_size):
for j in range(tim_size):
set_i = time_checkin_set[i]
set_j = time_checkin_set[j]
if len(set_i | set_j) != 0:
jaccard_ij = len(set_i & set_j) / len(set_i | set_j)
sim_matrix[i][j] = jaccard_ij
return sim_matrix
[docs]def parse_coordinate(coordinate):
items = coordinate[1:-1].split(',')
return float(items[0]), float(items[1])
[docs]def string2timestamp(strings, offset_frame):
ts = []
for t in strings:
dtstr = '-'.join([t[:4].decode(), t[4:6].decode(), t[6:8].decode()])
slot = int(t[8:]) - 1
ts.append(np.datetime64(dtstr, 'm') + slot * offset_frame)
return ts # [numpy.datetime64('2014-01-01T00:00'), ...]
[docs]def timestamp2array(timestamps, t):
"""
把时间戳的序列中的每一个时间戳转成特征数组，考虑了星期和小时，
时间戳: numpy.datetime64('2013-07-01T00:00:00.000000000')
Args:
timestamps: 时间戳序列
t: 一天有多少个时间步
Returns:
np.ndarray: 特征数组，shape: (len(timestamps), ext_dim)
"""
vec_wday = [time.strptime(
str(t)[:10], '%Y-%m-%d').tm_wday for t in timestamps]
vec_hour = [time.strptime(str(t)[11:13], '%H').tm_hour for t in timestamps]
vec_minu = [time.strptime(str(t)[14:16], '%M').tm_min for t in timestamps]
ret = []
for idx, wday in enumerate(vec_wday):
# day
v = [0 for _ in range(7)]
v[wday] = 1
if wday >= 5: # 0是周一, 6是周日
v.append(0) # weekend
else:
v.append(1) # weekday len(v)=8
# hour
v += [0 for _ in range(t)] # len(v)=8+T
hour = vec_hour[idx]
minu = vec_minu[idx]
# 24*60/T 表示一个时间步是多少分钟
# hour * 60 + minu 是从0:0开始到现在是多少分钟，相除计算是第几个时间步
# print(hour, minu, T, (hour * 60 + minu) / (24 * 60 / T))
v[int((hour * 60 + minu) / (24 * 60 / t))] = 1
# +8是因为v前边有表示星期的8位
if hour >= 18 or hour < 6:
v.append(0) # night
else:
v.append(1) # day
ret.append(v) # len(v)=7+1+T+1=T+9
return np.asarray(ret)
[docs]def timestamp2vec_origin(timestamps):
"""
把时间戳的序列中的每一个时间戳转成特征数组，只考虑星期，
时间戳: numpy.datetime64('2013-07-01T00:00:00.000000000')
Args:
timestamps: 时间戳序列
Returns:
np.ndarray: 特征数组，shape: (len(timestamps), 8)
"""
vec = [time.strptime(str(t)[:10], '%Y-%m-%d').tm_wday for t in timestamps]
ret = []
for i in vec:
v = [0 for _ in range(7)]
v[i] = 1
if i >= 5:
v.append(0) # weekend
else:
v.append(1) # weekday
ret.append(v)
return np.asarray(ret)