数据处理中的细节（自用）

# 同时满足两个条件
data = data.loc[(data.period!=-999)|(data.nuni!=-999)]

满足某条件
bid_nuni = bid_nuni[bid_nuni[‘nuni’]==1]

填充缺失值
data[‘nuni’] = data[‘nuni’].fillna(-999)

# 移除相同 样本
totalExposureLog = totalExposureLog.drop_duplicates(subset=[‘aid’,‘uid’,‘aid_location’,‘request_time’], keep=‘last’)#这里的last是取相似行的最后一行

# 移除pctr高于密集区的样本
totalExposureLog = totalExposureLog.loc[(totalExposureLog.pctr<=1000)]

# 移除quality_ecpm高于密集区的样本
totalExposureLog = totalExposureLog.loc[(totalExposureLog.quality_ecpm>=0)&(totalExposureLog.quality_ecpm<=80000)]

# 移除totalEcpm高于密集区的样本
totalExposureLog = totalExposureLog.loc[(totalExposureLog.totalEcpm<=120000)]

# 移除bid高于密集区的样本
totalExposureLog = totalExposureLog.loc[(totalExposureLog.bid<=15000)]

list转换为dataframe
crowd_feature = pd.DataFrame(crowd_data)

时间分割

def get_preprocessing(df_):
	    df = df_.copy()   
	    df['hour']  = df['create_order_time'].apply(lambda x:int(x[11:13]))
	    df['day']   = df['create_order_time'].apply(lambda x:int(x[8:10]))
	    df['month'] = df['create_order_time'].apply(lambda x:int(x[5:7]))
	    df['year']  = df['create_order_time'].apply(lambda x:int(x[0:4]))
	    df['date']  = (df['month'].values - 7) * 31 + df['day']    
	    del df['create_order_time']    
	    return df

train = get_preprocessing(train)
test  = get_preprocessing(test)

当缺失值比例占到40%或以上，建议删去这个特征，反而可以提高效果

在做一场比赛之前，一般都要先找几个相关的比赛熟悉一下套路

时间序列一定会用到滑窗