def get_mul_features(top_features):
print("===============正在进行特征乘法组合尝试========")
# 得到top7依次组合下的top15*7新乘法特征,每次均采用不同的样本
if(os.path.exists('./results/mul_features_top.kv')):
mul_features_top = joblib.load('./results/mul_features_top.kv')
else:
mul_features_top = {}
for x in top_features:
if x not in list(mul_features_top.keys()):
# x_train, scale_pos_weight = c_data_sample.get_sampled_data(c_data_sample.train_selected,
try:
c_data_sample.train_selected = c_data_sample.train_selected.drop(['target'], axis=1)
except:
print("drop train_selected.target")
pass # 0.7, 0.9)
x_train = c_data_sample.train_selected.copy()
scale_pos_weight = 4
print(x_train.head(1), b_feature_engineering.labels.head(1))
ytrain = pd.merge(x_train,
b_feature_engineering.labels, on='id').target
# print(ytrain.head())
for col in x_train.columns:
if col not in top_features and col != 'id':
x_train[col] = x_train[x] * x_train[col]
x_train.rename(
columns={col: (x+'_mul_'+col)}, inplace=True)
x_train = x_train.drop(top_features, axis=1)
clf = get_trained_xgb(x_train, ytrain.values, scale_pos_weight)
weight_kv = clf.get_booster().get_score()
mul_features_top[x] = list(weight_kv.keys())[:15]
del x_train
gc.collect()
joblib.dump(mul_features_top, './results/mul_features_top.kv')
print(mul_features_top[x])
# top内部组合
if ('top_mul_internal') not in list(mul_features_top.keys()):
print("mul 内部组合")
x_train = c_data_sample.train_selected.copy()
scale_pos_weight = 4
df = pd.DataFrame()
df['id'] = x_train.id
ytrain = pd.merge(df,
b_feature_engineering.labels, on='id').target
# print("特征维度",len(list(combinations(top_features,2))))
for (x, y) in list(combinations(top_features, 2)):
df[x+'_mul_'+y] = x_train[x] * x_train[y]
print(df.shape)
clf = get_trained_xgb(df, ytrain.values, scale_pos_weight)
weight_kv = clf.get_booster().get_score()
mul_features_top['top_mul_internal'] = list(weight_kv.keys())[:15]
del x_train
gc.collect()
joblib.dump(mul_features_top, './results/mul_features_top.kv')
def get_add_features(top_features):
# 得到top7依次组合下的top15*7新加法特征,每次均采用不同的样本
print("===============正在进行特征加法组合尝试========")
if(os.path.exists('./results/add_features_top.kv')):
add_features_top = joblib.load('./results/add_features_top.kv')
else:
add_features_top = {}
for x in top_features:
if x not in list(add_features_top.keys()):
x_train = c_data_sample.train_selected.copy()
scale_pos_weight = 4
ytrain = pd.merge(x_train,
b_feature_engineering.labels, on='id').target
for col in x_train.columns:
if col not in top_features and col != 'id':
x_train[col] = x_train[x] + x_train[col]
x_train.rename(
columns={col: (x+'_add_'+col)}, inplace=True)
x_train = x_train.drop(top_features, axis=1)
clf = get_trained_xgb(x_train, ytrain.values, scale_pos_weight)
weight_kv = clf.get_booster().get_score()
add_features_top[x] = list(weight_kv.keys())[:15]
del x_train
gc.collect()
joblib.dump(add_features_top, './results/add_features_top.kv')
print(add_features_top[x])
# top内部组合
if ('top_add_internal') not in list(add_features_top.keys()):
print("add 内部组合")
x_train = c_data_sample.train_selected.copy()
scale_pos_weight = 4
df = pd.DataFrame()
df['id'] = x_train.id
# print("特征维度",len(list(combinations(top_features,2))))
ytrain = pd.merge(df,
b_feature_engineering.labels, on='id').target
for (x, y) in list(combinations(top_features, 2)):
df[x+'_add_'+y] = x_train[x] + x_train[y]
print(df.shape)
clf = get_trained_xgb(df, ytrain.values, scale_pos_weight)
weight_kv = clf.get_booster().get_score()
add_features_top['top_add_internal'] = list(weight_kv.keys())[:15]
del x_train
gc.collect()
joblib.dump(add_features_top, './results/add_features_top.kv')
def apply_combine_features():
print("正在应用组合特征")
add_features_top = joblib.load('./results/add_features_top.kv')
mul_features_top = joblib.load('./results/mul_features_top.kv')
for v in add_features_top.values():
for col in v:
cols = str(col).split('_add_')
c_data_sample.train_selected[col] = c_data_sample.train_selected[cols[0]
]+c_data_sample.train_selected[cols[1]]
b_feature_engineering.test[col] = b_feature_engineering.test[cols[0]
]+b_feature_engineering.test[cols[1]]
for v in mul_features_top.values():
for col in v:
cols = str(col).split('_mul_')
c_data_sample.train_selected[col] = c_data_sample.train_selected[cols[0]
]*c_data_sample.train_selected[cols[1]]
b_feature_engineering.test[col] = b_feature_engineering.test[cols[0]
]*b_feature_engineering.test[cols[1]]
print("组合特征应用完毕")
print("特征排序：")
if not os.path.exists('./results/clf_1.model'):
# x_train, scale_pos_weight = c_data_sample.get_sampled_data(c_data_sample.train_selected,0.95, 0.9)
print("全部训练")
# clf = get_trained_xgb(x_train, scale_pos_weight)
x_train = c_data_sample.train_selected.copy()
scale_pos_weight = 4
ytrain = pd.merge(c_data_sample.train_selected,
b_feature_engineering.labels, on='id').target
clf = get_trained_xgb(c_data_sample.train_selected,
ytrain.values, scale_pos_weight)
# joblib.dump(clf,'./results/clf_1.model')
weight_kv = clf.get_booster().get_score()
# get_model_metrics(clf,x_train)
print(list(weight_kv.keys())[:30])
# X=c_data_sample.train_selected[~c_data_sample.train_selected.id.isin(
# list(x_train.id))]
# y=pd.merge(X, labels, on='id').target
# clf.fit(X.drop(['id'],axis=1),list(y),xgb_model =clf)
joblib.dump(clf, './results/clf_1.model')
# weight_kv = clf.get_booster().get_score()
# print(list(weight_kv.keys())[:30])
# joblib.dump(clf,'./results/feature_rank.kv')
print("特征排序结束")
y_predprob = clf.predict_proba(
b_feature_engineering.test.drop(['id'], axis=1))[:, 1]
df = pd.DataFrame()
df['id'] = b_feature_engineering.test.id
df['target'] = y_predprob
df.to_csv('./results/submission_2.csv', index=0, header=True)
# del x_train
# gc.collect()
else:
print("检测到已完成排序")
def get_mul_features(top_features):
# 得到top7依次组合下的top15*7新乘法特征,每次均采用不同的样本
if (os.path.exists('./results/mul_features_top.kv')):
mul_features_top = joblib.load('./results/mul_features_top.kv')
else:
mul_features_top = {}
for x in top_features:
if x not in list(mul_features_top.keys()):
x_train, scale_pos_weight = c_data_sample.get_sampled_data(
c_data_sample.train_selected, 0.6, 0.9)
for col in x_train.columns:
if col not in top_features and col != 'id':
x_train[col] = x_train[x] * x_train[col]
x_train.rename(columns={col: (x + '_mul_' + col)},
inplace=True)
x_train = x_train.drop(top_features, axis=1)
clf = get_trained_xgb(x_train, scale_pos_weight)
weight_kv = clf.get_booster().get_score()
mul_features_top[x] = list(weight_kv.keys())[:15]
del x_train
gc.collect()
joblib.dump(mul_features_top, './results/mul_features_top.kv')
print(mul_features_top[x])
# top内部组合
if ('top_mul_internal') not in list(mul_features_top.keys()):
print("mul 内部组合")
x_train, scale_pos_weight = c_data_sample.get_sampled_data(
c_data_sample.train_selected, 0.6, 0.9)
df = pd.DataFrame()
df['id'] = x_train.id
# print("特征维度",len(list(combinations(top_features,2))))
for (x, y) in list(combinations(top_features, 2)):
df[x + '_mul_' + y] = x_train[x] * x_train[y]
print(df.shape)
clf = get_trained_xgb(df, scale_pos_weight)
weight_kv = clf.get_booster().get_score()
mul_features_top['top_mul_internal'] = list(weight_kv.keys())[:15]
del x_train
gc.collect()
joblib.dump(mul_features_top, './results/mul_features_top.kv')
def main():
gc.collect()
a_preprocessing.get_preprocessed_data()
b_feature_engineering.make_features()
b_feature_engineering.one_hot_data()
eliminate_different_data()
# return
if (os.path.exists('./results/clf_0.model')):
clf_0 = joblib.load('./results/clf_0.model')
print("检测到初始模型已生成")
else:
# 1：0正负例样本采样，用于调节不平衡
x_train, scale_pos_weight = c_data_sample.get_sampled_data(
c_data_sample.train_selected, 0.95, 0.95) # 0.8，0.75建完测试时
# 在基本特征上得到第一个初始模型
clf_0 = get_trained_xgb(x_train, scale_pos_weight)
joblib.dump(clf_0, './results/clf_0.model')
get_model_metrics(clf_0, x_train)
del x_train
gc.collect()
# x_train, scale_pos_weight = c_data_sample.get_sampled_data(
# c_data_sample.train_selected, 0.9, 0.95)
# get_model_metrics(clf_0,x_train)
# del x_train
# gc.collect()
# 得到基础模型特征重要性排序
weight_kv_0 = clf_0.get_booster().get_score()
top_7 = list(weight_kv_0.keys())[:7]
print(list(weight_kv_0.keys())[:20])
# 得到top和其他列组合的加减法特征
get_mul_features(top_7)
get_add_features(top_7)
apply_combine_features()
def start(using_raw_features=False, nums=1200, scale_pos_weight_ctb=50, scale_pos_weight_xgb=8, neighbor_nums=50000):
gc.collect()
a_preprocessing.get_preprocessed_data() # 预处理
b_feature_engineering.make_features(
using_raw_features=using_raw_features) # 新建特征
b_feature_engineering.one_hot_data()
make_new_sample(b_feature_engineering.train,
b_feature_engineering.labels.target,
nums=nums) # 生成新样本数据
eliminate_different_data(neighbor_nums=neighbor_nums)
if(os.path.exists('./results/clf_0.model')):
clf_0 = joblib.load('./results/clf_0.model')
print("检测到初始模型已生成")
output(clf_0,b_feature_engineering.test,'./results/submission_xgb_1.csv')
else:
print("使用原始特征：{0},使用模型：{1},是否独热编码:{2},是否采用近邻点:{3},是否采用人工生成样本:{4}".format(
using_raw_features, 'ctb', 'yes', 'yes', 'yes'))
ytrain = pd.merge(c_data_sample.train_selected,
b_feature_engineering.labels, on='id').target
print(c_data_sample.train_selected.shape, len(ytrain))
ctb = get_trained_ctb(c_data_sample.train_selected,
ytrain.values, scale_pos_weight_ctb) # catboost结果
output(ctb,b_feature_engineering.test,'./results/submission_ctb_1.csv')
# 在基本特征上得到第一个初始模型
# print(c_data_sample.train_selected.head(5))
try:
c_data_sample.train_selected = c_data_sample.train_selected.drop(['target'], axis=1)
except:
print("drop train_selected.target")
pass
ytrain = pd.merge(c_data_sample.train_selected,
b_feature_engineering.labels, on='id').target
# print(c_data_sample.train_selected.head(5))
clf_0 = get_trained_xgb(
c_data_sample.train_selected, ytrain.values, scale_pos_weight_xgb)
output(clf_0, b_feature_engineering.test, './results/submission_xgb_1.csv')
# print(c_data_sample.train_selected.head(5))
joblib.dump(clf_0, './results/clf_0.model')
# gc.collect()
# x_train, scale_pos_weight = c_data_sample.get_sampled_data(
# c_data_sample.train_selected, 0.9, 0.95)
# get_model_metrics(clf_0,x_train)
# del x_train
# gc.collect()
# 得到基础模型特征重要性排序
weight_kv_0 = clf_0.get_booster().get_score()
top_7 = list(weight_kv_0.keys())[:7]
print(list(weight_kv_0.keys())[:20])
# 得到top和其他列组合的加减法特征
get_mul_features(top_7)
get_add_features(top_7)
apply_combine_features()