def SeqEncoder(df, pri_id, col, timesteps=5, encode_dim=5):
mode = df[[pri_id, col, 'day', 'time']]
# 获取每个用户的操作序列，按时间
# 按照天数降序排列，按照时间升序排列，然后按照UID聚集
mode.sort_values(by=['day', 'time'], ascending=[False, True], inplace=True)
temp = mode.groupby(pri_id)[col].apply(list).reset_index().rename(
columns={col: '%s_seq' % col})
temp['%s_seq_len' % col] = temp['%s_seq' % col].apply(lambda x: len(x))
ob_len = int(temp['%s_seq_len' % col].max() * 3.0 / 5.0)
# 改为timesteps的倍数
while ob_len % timesteps != 0:
ob_len += 1
# 不足长度的序列进行填充(选取长度为最长长度的5/3)
def Padding(x):
if len(x) < ob_len:
# 填充(Nan)
x += ['Nan' for i in range(ob_len - len(x))]
else:
# 截取
x = x[:ob_len]
return x
temp['%s_seq' % col] = temp['%s_seq' % col].apply(Padding)
# 编码
from sklearn.preprocessing import LabelEncoder
le = LabelEncoder()
mode['encoder'] = le.fit_transform(mode[col])
# 建立字典
seq_dic = {}
encoder = mode[[col, 'encoder']].drop_duplicates()
for key, val in zip(encoder[col], encoder['encoder']):
seq_dic[key] = val
seq_dic['Nan'] = mode['encoder'].max() + 1
# 对序列进行转化
temp['%s_seq' % col] = temp['%s_seq' %
col].apply(lambda x: [seq_dic[i] for i in x])
matrix = np.asarray([i for i in temp['%s_seq' % col]])
# 全连接层编码
# res = _F.AutoEncoder(matrix,encode_dim)
# LSTM编码
res = _F.AutoEncoderLSTM(matrix, 5, timesteps)
# 保存结果
results = pd.DataFrame()
results[pri_id] = temp[pri_id]
for i in range(encode_dim):
results[col + "_encode_" + str(i)] = res[:, i]
results.to_csv(data_path + "data/_F_seq_encode_%s.csv" % col, index=False)
return results
