开源一个0.827的baseline
没做太多特征，读数据，看分布，如果分布是长尾分布就加个变换
去掉相关系数低于0.05的特征
对某些在某些区间聚集较为明显的特征分桶处理
网格调参，我还没跳到最优，太慢了
采用xgb，rf融合模型
注释已经很详细了
进不去前14，拿不了复赛名额，就开源吧
是用jupyter写的，ipynb文件发到了大赛群里
#!/usr/bin/env python# coding: utf-8# In[1]:import warningswarnings.filterwarnings("ignore")import numpy as npimport pandas as pdimport matplotlib.pyplot as pltimport seaborn as sns# ## 读取数据# In[2]:train_set = pd.read_csv('./train_set.csv')train_set.head()# ### 正样本占比0.2# In[3]:train_label = pd.read_csv('./train_label.csv')train_label[train_label['label'] == 1].shape[0] / train_label.shape[0]# In[4]:# 测试集读入test = pd.read_csv('result_predict_A.csv')test['label'] = -1# test.info()# In[5]:# 全样本构建,flag判断是训练集还是测试集train = pd.merge(train_set, train_label)all_data = train.append(test).reset_index(drop='True')all_data['flag'] = all_data['label'].map(lambda x: 'test' if x == -1 else 'train')# all_data['X5'].mode()all_data['X5'] = all_data['X5'].fillna('大众用户')all_data[all_data['X6'].isnull() & (all_data['label'] == -1)]all_data.head()# In[91]:all_data.info()# ### 测试集中X6到17缺失的人直接赋值0# ### X5用众数即大众用户填充# ### X6,X7,X8具有强相关性,X4和userid有强相关性，X3，32，33和6，7，8有关系# In[6]:train.corr() #3,6,7,8,32,33,24# In[7]:corr_dict = dict(train.corr()[train.corr() > 0.1].iloc[:, -1].dropna())columns = list(corr_dict.keys())[:-1] # 强相关的列# ### 进一步分析相关性# In[8]:train['X5'] = train['X5'].fillna('大众用户')set(train['X5'].to_list())# ## 特征工程# ### 考虑6-14要不要删除# ### 24,28极强相关# In[10]:# columns = ['X3', 'X5', 'X15', 'X16', 'X17', 'X24', 'X29', 'X32', 'X34', 'X37', 'X39']# columns = ['X' + str(i) for i in [3,5,6,7,8,9,10,11,12,13,14,15,16,17,24,29,32,33,34,37,39]]columns = ['X' + str(i) for i in [3,5,6,7,8,9,12,15,16,17,24,29,32,34,37,38,39,41,42,43]]columns.append('user_id')# ### 尝试加入其他特征# In[11]:all_data = all_data[[i for i in columns] + ['label', 'flag']]all_data['X38'] = all_data['X38'].fillna(0)all_data.head()# In[12]:all_data = all_data.dropna(axis=0, subset=['X16'])all_data.info()# 改变数据分布# In[13]:all_data['X8'] = np.log(all_data['X8'].values+1)sns.kdeplot(all_data['X8'], color="Red", shade = True)# In[14]:all_data['X7'] = np.log(all_data['X7'].values+1)sns.kdeplot(all_data['X7'], color="Red", shade = True)# In[15]:all_data['X6'] = np.log(all_data['X6'].values+1)sns.kdeplot(all_data['X6'], color="Red", shade = True)# In[16]:all_data['X9'] = np.log(all_data['X9'].values+1)sns.kdeplot(all_data['X9'], color="Red", shade = True)# In[17]:all_data['X15'] = np.log(all_data['X15'].values+1)sns.kdeplot(all_data['X15'], color="Red", shade = True)# In[18]:all_data['X16'] = np.log(all_data['X16'].values+1)sns.kdeplot(all_data['X16'], color="Red", shade = True)# 新增特征# In[19]:def trans(x):if x <= 1:return 0elif x > 1 and x < 6:return 1else:return 2# In[20]:all_data['X16_range'] = all_data['X16'].apply(trans)# In[21]:all_data['X17'] = np.log(all_data['X17'].values+1)sns.kdeplot(all_data['X17'], color="Red", shade = True)# In[22]:sns.kdeplot(all_data['X24'], color="Red", shade = True)# In[23]:sns.kdeplot(all_data['X29'], color="Red", shade = True)# In[24]:all_data.head()# ### 用数字特征填补缺失值# In[25]:sns.kdeplot(all_data['X3'], color="Red", shade = True)all_data['X3'] = all_data['X3'].fillna(3)# In[26]:sns.kdeplot(all_data['X29'], color="Red", shade = True) # 24,32，33all_data['X29'] = all_data['X29'].fillna(0)# In[35]:sns.kdeplot(all_data['X34'], color="Red", shade = True) # 24,32，33all_data['X34'] = all_data['X34'].fillna(0)# In[27]:all_data = pd.concat([pd.get_dummies(all_data['X5']), all_data], axis=1).drop('X5', axis=1)all_data.head()# 处理X32# In[31]:all_data['X32'] = np.log(all_data['X32'].values+1)sns.kdeplot(all_data['X32'], color="Red", shade = True)# In[ ]:# 填充缺失值from sklearn.ensemble import RandomForestRegressortemp = all_data# X32known = temp[temp['X32'].notnull()]unknown = temp[temp['X32'].isnull()]X = known.drop(['user_id', 'X32', 'label', 'flag'], axis=1).valuesy = known['X32'].valuesrfr = RandomForestRegressor(random_state=0, n_estimators=100)rfr.fit(X, y)predict_X32 = rfr.predict(unknown.drop(['user_id', 'X32', 'label', 'flag'], axis=1).values)all_data.loc[all_data['X32'].isnull(), 'X32'] = predict_X32# 新增特征# In[124]:def transX32(x):if x < 2.7:return 0elif 2.7 <= x < 3.15:return 1elif 3.15 <= x < 3.92:return 2elif 3.92 <= x < 4.4:return 3elif 4.4 <= x <4.9:return 3else:return 4# In[125]:all_data['X32_range'] = all_data['X32'].apply(transX32)# In[128]:del all_data['X38']# ### 分割训练集，验证集，测试集# In[137]:train = all_data[all_data['flag'] == 'train'].drop(['flag', 'user_id'], axis=1)test = all_data[all_data['flag'] == 'test'].drop(['label', 'flag', 'user_id'], axis=1).reset_index(drop=True)# In[140]:import xgboost as xgbfrom tqdm import tqdmfrom xgboost.sklearn import XGBClassifierfrom sklearn.model_selection import GridSearchCVfrom sklearn.ensemble import RandomForestClassifierfrom sklearn.model_selection import train_test_splitfrom sklearn.metrics import mean_squared_error, f1_score # 均方误差import catboost as cbX_train, X_cv, y_train, y_cv = train_test_split(train.drop(['label'], axis=1), train['label'], test_size=0.2)# ### 网格调参# In[758]:param_grid = [{'n_estimators': list(range(100, 501, 100)), 'max_depth':list(range(2,21,5))}]rf = RandomForestClassifier()grid_search_rf = GridSearchCV(rf, param_grid, cv=5,scoring='f1')grid_search_rf.fit(train.drop(['label'], axis=1), train['label'])print(grid_search_rf.best_estimator_) # max_depth=17,n_estimators=500# In[760]:param_dist = {'n_estimators':list(range(20, 141, 20)), # 120'max_depth':list(range(2,15,5)), # 7'learning_rate':list(np.linspace(0.01,2,5)), # 0.01# 'subsample':list(np.linspace(0.7,0.9,5)),# 'colsample_bytree':list(np.linspace(0.5,0.98,3)),# 'min_child_weight':list(range(1,9,3)) # 6}xgb = XGBClassifier()grid_search_xgb = GridSearchCV(xgb, param_dist,cv = 3,n_jobs = -1, scoring='f1')grid_search_xgb.fit(train.iloc[:, :-1], train['label'])print(grid_search_xgb.best_estimator_)# ### 验证集效果# In[141]:train = all_data[all_data['flag'] == 'train'].drop(['flag', 'user_id'], axis=1)test = all_data[all_data['flag'] == 'test'].drop(['label', 'flag', 'user_id'], axis=1).reset_index(drop=True)X_train, X_cv, y_train, y_cv = train_test_split(train.drop(['label'], axis=1), train['label'], test_size=0.2)# In[142]:rf = RandomForestClassifier(n_estimators=500, max_depth=17).fit(X_train, y_train)print('rf F1: {}' .format(f1_score(rf.predict(X_cv), y_cv)))# In[143]:xgb = XGBClassifier().fit(X_train, y_train)print('xgb F1: {}' .format(f1_score(xgb.predict(X_cv), y_cv)))# In[168]:from sklearn.linear_model import LogisticRegressionfor x in np.linspace(500, 1500, 10):clf3 = LogisticRegression(penalty='l2', C=0.1, max_iter=x, tol=1e-4, solver='lbfgs').fit(X_train, y_train)print(x)print('lr F1: {}' .format(f1_score(clf3.predict(X_cv), y_cv)))# In[203]:clf4 = cb.CatBoostClassifier(n_estimators=7000).fit(X_train, y_train)print('catboost F1: {}' .format(f1_score(clf4.predict(X_cv), y_cv)))# In[206]:y_pred_1 = rf.predict_proba(X_cv)[:, 0]y_pred_2 = clf4.predict_proba(X_cv)[:, 0]y_pred = (y_pred_1 + y_pred_2 ) / 2y_pred = list(map(lambda x: 1 if x<0.62 else 0, y_pred))print(f1_score(y_pred, y_cv))# 遍历找到阈值score_lst = []for i in list(np.linspace(0.45,0.75,100)):i = round(i, 4)y_pred = (y_pred_1 + y_pred_2 ) / 2y_pred = list(map(lambda x: 1 if x<i else 0, y_pred))score = f1_score(y_pred, y_cv)score_lst.append([i, score])print('i={}, total F1: {}' .format(i, score))score_lst = np.array(score_lst)plt.plot(score_lst[:, 0], score_lst[:, 1])# ### 预测# In[1]:clf1 = RandomForestClassifier(n_estimators=500, max_depth=17)clf2 = cb.CatBoostClassifier(n_estimators=5000)clf1.fit(train.drop(['label'], axis=1), train['label'])print('训练完了')clf2.fit(train.drop(['label'], axis=1), train['label'])# In[232]:y_pred_1 = clf1.predict_proba(test)[:, 0]y_pred_2 = clf2.predict_proba(test)[:, 0]y_pred = (y_pred_1 + y_pred_2 ) / 2 y_pred = list(map(lambda x: 1 if x<0.75 else 0, y_pred))# 添加特殊用户# In[233]:temp = pd.read_csv('result_predict_A.csv')temp[temp['X16'].isnull()]# In[190]:extra = pd.DataFrame([['2697592699877', 0], ['2697527496793', 0], ['2697624945417', 0]], columns=['user_id', 'label'])extra.head()# In[234]:result = pd.read_csv('result_predict_A.csv')result = result.dropna(axis=0, subset=['X16'])result['label'] = y_predresult = result[['user_id', 'label']]# 加入X16为NAN的三个样本result = result.append(extra)result.head()# In[235]:result.shape# In[236]:result.to_csv('./submit.csv', index=False)
