DataFrame衍生新特征操作

1.DataFrame中某一列的值衍生为新的特征

#将LBL1特征的值衍生为one-hot形式的新特征
piao=df_train_log.LBL1.value_counts().index
#先构造一个临时的df
df_tmp=pd.DataFrame({'USRID':df_train_log.drop_duplicates('USRID').USRID.values})
#将所有的新特征列都置为0
for i in piao:
    df_tmp['PIAO_'+i]=0

#进行分组便利，有这个特征就置为1，原数据每个USRID有多条记录，所以分组统计
group=df_train_log.groupby(['USRID'])
for k in group.groups.keys():
    t = group.get_group(k)
    id=t.USRID.value_counts().index[0]
    tmp_list=t.LBL1.value_counts().index
    for j in tmp_list:
        df_tmp['PIAO_'+j].loc[df_tmp.USRID==id]=1

 2.分组统计，选出同一USRID下该变量中出现次数最多的值项

group=df_train_log.groupby(['USRID'])
lt=[]
list_max_lbl1=[]
list_max_lbl2=[]
list_max_lbl3=[]
for k in group.groups.keys():
    t = group.get_group(k)
    #通过value_counts找出出现次数最多的项
    argmx = np.argmax(t['EVT_LBL'].value_counts())
    lbl1_max=np.argmax(t['LBL1'].value_counts())
    lbl2_max=np.argmax(t['LBL2'].value_counts())
    lbl3_max=np.argmax(t['LBL3'].value_counts())
    list_max_lbl1.append(lbl1_max)
    list_max_lbl2.append(lbl2_max)
    list_max_lbl3.append(lbl3_max)
    #只留下出现次数最多的项
    c = t[t['EVT_LBL']==argmx].drop_duplicates('EVT_LBL')
    #放入list中
    lt.append(c)
#构造一个新的df
df_train_log_new = pd.concat(lt)
#另外又构造了三个特征，LBL1-LBL3分别出现次数最多的项
df_train_log_new['LBL1_MAX']=list_max_lbl1
df_train_log_new['LBL2_MAX']=list_max_lbl2
df_train_log_new['LBL3_MAX']=list_max_lbl3

 3.衍生出某天是否发生的ont-hot新特征

#创造临时df，星期三，星期六，星期七，都默认置为0
df_day=pd.DataFrame({'USRID':df_train_log.drop_duplicates('USRID').USRID.values})
df_day['weekday_3']=0
df_day['weekday_6']=0
df_day['weekday_7']=0

#分组统计，有就置为1，没有置为0
group=df_train_log.groupby(['USRID'])
for k in group.groups.keys():
    t = group.get_group(k)
    id=t.USRID.value_counts().index[0]
    tmp_list=t.occ_dayofweek.value_counts().index
    for j in tmp_list:
        if j==3:
            df_day['weekday_3'].loc[df_tmp.USRID==id]=1
        elif j==6:
            df_day['weekday_6'].loc[df_tmp.USRID==id]=1
        elif j==7:
            df_day['weekday_7'].loc[df_tmp.USRID==id]=1

 

4.查看用户一共停留在APP上多少秒，共有几天看了APP

#首先将日期转化为时间戳，并赋予一个新特征
tmp_list=[]
for i in df_train_log.OCC_TIM:
    d=datetime.datetime.strptime(str(i),"%Y-%m-%d %H:%M:%S")
    evt_time = time.mktime(d.timetuple())
    tmp_list.append(evt_time)

df_train_log['time']=tmp_list

#每下一行减去上一行，得到app停留时间
df_train_log['diff_time']=df_train_log.time-df_train_log.time.shift(1)

#构造一个新的dataFrame，分组得到查看app的天数
df_time=pd.DataFrame({'USRID':df_train_log.drop_duplicates('USRID').USRID.values})
#有几天查看
df_time['days']=0
group=df_train_log.groupby(['USRID'])
for k in group.groups.keys():
    t = group.get_group(k)
    id=set(t.USRID).pop()
    df_time['days'].loc[df_time.USRID==id]= len(t.occ_day.value_counts().index)

#去掉一些异常时间戳，比如间隔两天的相减，肯定不合适，na的也去掉了
df_train_log=df_train_log[(df_train_log.diff_time>0)&(df_train_log.diff_time<8000)]
#累计停留时间
group_stayTime=df_train_log['diff_time'].groupby(df_train_log['USRID']).sum()
#创造新的df
df_tmp=pd.DataFrame({'USRID':list(group_stayTime.index.values),'stay_time':list(group_stayTime.values)})

#合并成一个新的df
df=pd.merge(df_time,df_tmp,on=['USRID'],how='left')
#合并后，缺失的停留时间，置为0
df.fillna(0,axis=1,inplace=True)