机器学习之基于xgboost的特征筛选
本⽂主要是基于xgboost进⾏特征选择，很多⼈都知道在后⾯的模型选择时，xgboost模型是⼀个⾮常热门的模型。但其实在前⾯特征选择部分，基于xgboost进⾏特征筛选也⼤有可为。
#coding=utf-8
import pandas as pd
import xgboost as xgb
import os,random,pickle
os.mkdir('featurescore')
train = pd.read_csv('../../data/train/train_x_rank.csv')
train_target = pd.read_csv('../../data/train/train_master.csv',encoding='gb18030')[['Idx','target']]
train = pd.merge(train,train_target,on='Idx')
train_y = train.target
train_x = train.drop(['Idx','target'],axis=1)
dtrain = xgb.DMatrix(train_x, label=train_y)
test = pd.read_csv('../../data/test/test_x_rank.csv')
test_Idx = test.Idx
test = test.drop('Idx',axis=1)
dtest = xgb.DMatrix(test)
train_test = pd.concat([train,test])
_csv('rank_feature.csv',index=None)
print print(train_test.shape)
"""
params={
'booster':'gbtree',
'objective': 'rank:pairwi',
'scale_pos_weight': float(len(train_y)-sum(train_y))/float(sum(train_y)),
'eval_metric': 'auc',
'gamma':0.1,
'max_depth':6,
修改文件时间'lambda':500,
'subsample':0.6,
'colsample_bytree':0.3,
'min_child_weight':0.2,
'eta': 0.04,
'ed':1024,
'nthread':8
}
xgb.cv(params,dtrain,num_boost_round=1100,nfold=10,metrics='auc',show_progress=3,ed=1024)#733
"""
def pipeline(iteration,random_ed,gamma,max_depth,lambd,subsample,colsample_bytree,min_child_weight):
叠小纸船params={
'booster':'gbtree',
'objective': 'rank:pairwi',
'scale_pos_weight': float(len(train_y)-sum(train_y))/float(sum(train_y)),
'eval_metric': 'auc',
'gamma':gamma,
'max_depth':max_depth,
'lambda':lambd,
'subsample':subsample,
'colsample_bytree':colsample_bytree,
'colsample_bytree':colsample_bytree,
'min_child_weight':min_child_weight,
'eta': 0.2,
欢子的歌曲
'ed':random_ed,
'nthread':8
}
watchlist = [(dtrain,'train')]
model = ain(params,dtrain,num_boost_round=700,evals=watchlist)
#model.save_model('./model/xgb{0}.model'.format(iteration))
#predict test t
#test_y = model.predict(dtest)
#test_result = pd.DataFrame(test_Idx,columns=["Idx"])
#test_result["score"] = test_y
#_csv("./preds/xgb{0}.csv".format(iteration),index=None,encoding='utf-8')
#save feature score
feature_score = _fscore()
feature_score = sorted(feature_score.items(), key=lambda x:x[1],rever=True)
fs = []
for (key,value) in feature_score:
fs.append("{0},{1}\n".format(key,value))
石将军石勇with open('./featurescore/feature_score_{0}.csv'.format(iteration),'w') as f:
f.writelines("feature,score\n")
f.writelines(fs)
if __name__ == "__main__":
random_ed = range(10000,20000,100)
gamma = [i/1000.0 for i in range(0,300,3)]
儿童挑食max_depth = [5,6,7]
lambd = range(400,600,2)
subsample = [i/1000.0 for i in range(500,700,2)]
colsample_bytree = [i/1000.0 for i in range(550,750,4)]
min_child_weight = [i/1000.0 for i in range(250,550,3)]
random.shuffle(random_ed)
random.shuffle(gamma)
random.shuffle(max_depth)
random.shuffle(lambd)
random.shuffle(subsample)
random.shuffle(colsample_bytree)
微信红包秒抢random.shuffle(min_child_weight)
with open('params.pkl','w') as f:
pickle.dump((random_ed,gamma,max_depth,lambd,subsample,colsample_bytree,min_child_weight),f)
for i in range(36):
pipeline(i,random_ed[i],gamma[i],max_depth[i%3],lambd[i],subsample[i],colsample_bytree[i],min_child_weight[i])
因为xgboost的参数选择⾮常重要，因此进⾏了参数shuffle的操作。最后可以基于以上不同参数组合的xgboost所得到的feature和socre，再进⾏score平均操作，筛选出⾼得分的特征。
import pandas as pd
import os
files = os.listdir('featurescore')
fs = {}
for f in files:
t = pd.read_csv('featurescore/'+f)
t.index = t.feature
t = t.drop(['feature'],axis=1)
d = t.to_dict()['score']
for key in d:
if fs.has_key(key):
fs[key] += d[key]
el:
fs[key] = d[key]
韩国海带汤>牧云陆
fs = sorted(fs.items(), key=lambda x:x[1],rever=True)
t = []
for (key,value) in fs:
t.append("{0},{1}\n".format(key,value))
with open('rank_feature_score.csv','w') as f:
f.writelines("feature,score\n")
f.writelines(t)
这⾥得出了每个特征的总分，每个都除以36就是平均分了。最后按照平均分取出topn就可以。我的理解是这样⼦。然后觉得这种⽅法太耗时了。
