def combine_backward(df, order, period=30, strict=[]):
"""
combine_backward(df, order, period=30, strict=[]):
将一组信号按order中的顺序进行组合，在order[i+1]序列的两个1的区间内，匹配order[i]中最靠前的1。
即时间靠前的下层信号会覆盖后面的下层信号。
得到一个集合，集合中每条记录代表一个可行的信号组合。
Input:
df: (DataFrame): 一组信号序列，至少包含order中涉及的列
order: (list of string): 按序排好的一组信号的名称，至少两列
period: (int): 从末位信号开始算，一组信号的最长周期。超过该周期仍未凑够一组信号则放弃该条记录。
strict: (list of boolean): 长度为len(order) - 1，strict[i]指示order[i]的信号出现时间是否必须早于order[i+1]的时间。
True必须严格早于；False可以同一天连续触发。
参数默认值为[True]*(len(order)-1)，若传入列表长度不足，则尾部用True填充
Output:
(DataFrame[order]):返回DataFrame，每条记录（一行）代表一个可行的信号组合，每列中的值为该信号在df中的索引
"""
rdf = pd.DataFrame()
#将df最后一列所有信号的位置全部放进rdf中
rdf[order[-1]] = pd.Series(np.where(df[order[-1]] == 1)[0])
if len(order) > 1:
if len(strict) < len(order) - 1:
rest_len = len(order) - 1 - len(strict)
strict.extend([True] * rest_len)
last_col = rdf[order[-1]]
#每个信号的范围只与last_col-period有关，前面信号不对后面信号的范围造成影响
if period is None:
range_sr = pd.Series([(0, last_col[i]) if i == 0 \
else (last_col[i-1], last_col[i]) \
for i in range(len(rdf))])
else:
range_sr = pd.Series([(max(0,last_col[i]-period),last_col[i]) if i == 0\
else (last_col[i]-period, last_col[i])\
for i in range(len(rdf))])
odr = order.copy()
odr.reverse()
st = strict.copy()
st.reverse()
#对每一列信号操作
for i in range(1, len(odr)):
#找该行的信号位置
signal_loc = np.where(df[odr[i]] == 1)[0]
tmp_sr = range_sr.map(lambda x: _backward(x, signal_loc, st[i-1]), \
na_action='ignore')
rdf[odr[i]] = tmp_sr.map(lambda x: x[1])
range_sr = tmp_sr
#去掉rdf中包含np.NAN的记录（行）
if rdf is not None and len(rdf) > 0:
rdf = rdf.drop(np.where(np.isnan(rdf))[0])
rdf = rdf[order].reset_index(drop=True)
else:
rdf = None
return rdf
def _backward(x, signal_loc, strict):
#若strict=False即宽松条件，当进入的tuple为(0,0)时，需要让left为-1才能套公式
left = -1 if (x[0] == 0 and not strict) else x[0]
right = x[1]
if strict:
loc = np.where((signal_loc >= left) & (signal_loc < right))[0]
else:
loc = np.where((signal_loc >= left) & (signal_loc <= right))[0]
right = signal_loc[loc.max()] if len(loc) > 0 else np.NAN
result = (left, right) if (not np.isnan(right)) else (np.NAN, np.NAN)
return result
def ix_nan(self,colnames=None,indices=None):
# get the indices based on input.
indices=self.getindices(indices)
# get the column names over which to iterate
colnames=self.getcolnames(colnames)
for colname in colnames:
#print('XXX',indices)
(null,) = np.where(pd.isnan(self.t.loc[indices,colname]))
indices = indices[null]
#print('YYY',notnull)
return(indices)
'long_term_incentive', 'restricted_stock', 'director_fees', 'to_messages', \
'from_messages', 'norm_from_poi_to_this_person', 'norm_from_this_person_to_poi', \
'norm_shared_receipt_with_poi', 'from_poi_to_this_person', 'from_this_person_to_poi', 'shared_receipt_with_poi']
figure = plt.figure(figsize=(27, 20))
j = 1
for i, feature1 in enumerate(features_list2[1:]):
y_name = features_list2[0]
key = feature1
bin_step = 20
all_data = df[[y_name, key]]
# Remove NaN values from Age data
all_data = all_data[~np.isnan(all_data[key])]
# Divide the range of data into bins and count survival rates
min_value = all_data[key].min()
max_value = all_data[key].max()
value_range = max_value - min_value
bins = np.arange(min_value, max_value + value_range/bin_step, value_range/bin_step )
y0 = all_data[all_data[y_name] == 0][key].reset_index(drop = True)
y1 = all_data[all_data[y_name] == 1][key].reset_index(drop = True)
ax = plt.subplot(len(features_list2)/4+1, 4, j)
ax.hist(y0, bins = bins, alpha = 0.6, color = 'red', label = 'y0')
ax.hist(y1, bins = bins, alpha = 0.6, color = 'green', label = 'y1')
ax.set_xlim(bins.min(), bins.max())
ax.set_title(key)
# ax.legend(framealpha = 0.8)
j+=1