CCF2020 大資料時代的Serverless作業負載預測賽道baseline：(0.211+)

附一下官網鏈接：https://www.datafountain.cn/competitions/468

之前剛開賽的時候,寫了個baseline,分享一下，目前排名有點落后,大家隨便看看，

import pandas as pd
from datetime import datetime
import time
from sklearn.preprocessing import LabelEncoder

def make_label(data):

	'''構建十個目標 當做多目標任務做 
	沒有用多目標預測的庫 而是分開預測的 
	這里應該能夠上點分'''
	
    data['CPU_USAGE_1']=data.CPU_USAGE.shift(-1)
    data['CPU_USAGE_2']=data.CPU_USAGE.shift(-2)
    data['CPU_USAGE_3']=data.CPU_USAGE.shift(-3)
    data['CPU_USAGE_4']=data.CPU_USAGE.shift(-4)
    data['CPU_USAGE_5']=data.CPU_USAGE.shift(-5)
    
    data['LAUNCHING_JOB_NUMS_1']=data.LAUNCHING_JOB_NUMS.shift(-1)
    data['LAUNCHING_JOB_NUMS_2']=data.LAUNCHING_JOB_NUMS.shift(-2)
    data['LAUNCHING_JOB_NUMS_3']=data.LAUNCHING_JOB_NUMS.shift(-3)
    data['LAUNCHING_JOB_NUMS_4']=data.LAUNCHING_JOB_NUMS.shift(-4)
    data['LAUNCHING_JOB_NUMS_5']=data.LAUNCHING_JOB_NUMS.shift(-5)
    
    '''
    因為使用shift 所有會產生一些缺失值 這里直接刪了 
    數量不多 應該問題不大
    '''
    return data.dropna()
    

def process(df):
	'''
	對時間進行了一點處理
	訓練集沒有問題
	測驗集出現了2023年時間
	應該是官方脫敏的原因
	'''
    df.DOTTING_TIME/=1000
    df.DOTTING_TIME=list(map(lambda x:time.strftime("%Y-%m-%d %H:%M:%S", time.localtime(x)),df.DOTTING_TIME))
    df=df.sort_values(['QUEUE_ID','DOTTING_TIME'])
    df['DOTTING_TIME']=pd.to_datetime(df.DOTTING_TIME)
    return df

train=pd.read_csv('train.csv')
test=pd.read_csv('evaluation_public.csv')
train=process(train)
test=process(test)
#將之后五個時間點的數值作為label
train=train.groupby('QUEUE_ID').apply(make_label)

train=train.reset_index(drop=True)

#特征值轉化
encode_STATUS=LabelEncoder()
encode_QUEUE_TYPE=LabelEncoder()
encode_PLATFORM=LabelEncoder()
encode_RESOURCE_TYPE=LabelEncoder()

train.STATUS=encode_STATUS.fit_transform(train.STATUS)
test.STATUS=encode_STATUS.transform(test.STATUS)

train.QUEUE_TYPE=encode_QUEUE_TYPE.fit_transform(train.QUEUE_TYPE)
test.QUEUE_TYPE=encode_QUEUE_TYPE.transform(test.QUEUE_TYPE)

train.PLATFORM=encode_PLATFORM.fit_transform(train.PLATFORM)
test.PLATFORM=encode_PLATFORM.transform(test.PLATFORM)

train.RESOURCE_TYPE=encode_RESOURCE_TYPE.fit_transform(train.RESOURCE_TYPE)
test.RESOURCE_TYPE=encode_RESOURCE_TYPE.transform(test.RESOURCE_TYPE)

train.drop(['DOTTING_TIME'],axis=1,inplace=True)
test.drop(['DOTTING_TIME'],axis=1,inplace=True)

targets_names=['CPU_USAGE_1','LAUNCHING_JOB_NUMS_1','CPU_USAGE_2','LAUNCHING_JOB_NUMS_2','CPU_USAGE_3','LAUNCHING_JOB_NUMS_3',
'CPU_USAGE_4','LAUNCHING_JOB_NUMS_4','CPU_USAGE_5','LAUNCHING_JOB_NUMS_5']
targets=[]
for i in targets_names:
    targets.append(train[i])
    train.drop(i,axis=1,inplace=True)

'''
只保留測驗集最后一條樣本進行預測,測驗集利用率還是比較低
'''
test=test.drop_duplicates(subset=['ID'],keep='last')
test_id=test.ID
test.drop('ID',axis=1,inplace=True)

from catboost import CatBoostRegressor
from lightgbm import LGBMRegressor
from sklearn.model_selection import train_test_split

df=pd.DataFrame()
df['ID']=test_id
model=LGBMRegressor(n_estimators=100000,eval_metric='mse')
for i in targets:
    train_x,test_x,train_y,test_y=train_test_split(train,i,test_size=0.2,random_state=42)
    model.fit(train_x,train_y,eval_set=(test_x,test_y),early_stopping_rounds=50,verbose=100)
    df[i.name]=model.predict(test,num_iteration=model.best_iteration_)

為了防止號沒了,改動了一下劃分訓練集測驗集的隨機種子,可能還有一些庫的版本差距,差距應該不大,比較慫,希望沒有直接提交的,萬一改了個隨機種子直接上分就有點尷尬(手動狗頭)