和鲸社区AI实训季 客户购买预测

1. 赛题

本练习赛与葡萄牙银行机构的营销活动相关。这些营销活动一般以电话为基础，银行的客服人员至少联系客户一次，以确认客户是否有意愿购买该银行的产品（定期存款）。

任务是基本类型为分类任务，即预测客户是否购买该银行的产品。

2.详细步骤与代码实现

2.1 数据加载与合并

import pandas as pd
import numpy as np
from sklearn.model_selection import KFold
from catboost import CatBoostClassifier

train = pd.read_csv('/home/mw/input/purchase_pred8546/purchase_prediction_train_set.csv')
test = pd.read_csv('/home/mw/input/purchase_pred8546/purchase_prediction_test_set.csv')

# 合并处理
data = pd.concat([train, test], ignore_index=True, sort=False)

2.2 特征工程构建

特征工程分类：

# 处理 pdays：999表示缺失值
data['pdays_missing'] = (data['pdays'] == 999).astype(int)
data.loc[data['pdays'] == 999, 'pdays'] = -1

# balance 是否为负 - 财务健康度指标
data['balance_negative'] = (data['balance'] < 0).astype(int)

# duration 强化特征 - 交流时长相关
data['duration_log'] = np.log1p(data['duration'])  
data['dur_prev'] = data['duration'] * (data['previous'] + 1) 
data['dur_per_call'] = data['duration'] / (data['campaign'] + 1) 
data['pdays_recip'] = 1 / (data['pdays'] + 2)  

# poutcome 是否成功 × duration - 成功历史与当前交流的交互
data['dur_success'] = data['duration'] * (data['poutcome'] == 'success').astype(int)

# month 数值编码 - 捕捉季节性趋势
data['month_num'] = data['month'].map({
    'jan':1,'feb':2,'mar':3,'apr':4,'may':5,'jun':6,
    'jul':7,'aug':8,'sep':9,'oct':10,'nov':11,'dec':12
})

2.3：数据拆分与预处理

• 重新拆分数据集，准备建模数据

• 根据目标变量y的存在与否拆分数据

• 移除ID列（非特征变量）

• 明确指定类别特征，CatBoost会自动处理

  train_processed = data[data['y'].notna()].copy() test_processed = data[data['y'].isna()].copy()

  特征和目标分离

  X = train_processed.drop(['y','ID'], axis=1) y = train_processed['y'].astype(int) X_test = test_processed.drop(['y','ID'], axis=1)

  类别特征定义

  cat_cols = [ 'job','marital','education','default','housing','loan', 'contact','month','poutcome' ]

2.4：模型参数配置

目的：设置CatBoost最优参数组合

参数调优重点：

• bootstrap_type='Bayesian'：关键改进，提升稳定性

• iterations=3000 + learning_rate=0.02：小步长多树策略

• od_type='Iter'：自动过拟合检测

  model = CatBoostClassifier( iterations=3000,
  learning_rate=0.02,
  depth=7,
  l2_leaf_reg=3,
  random_strength=1.5,
  bagging_temperature=0.6,
  border_count=254,
  loss_function='Logloss',
  eval_metric='AUC',
  bootstrap_type='Bayesian', od_type='Iter',
  od_wait=80,
  random_seed=42,
  task_type='CPU'
  )

2.5：交叉验证训练

使用5折交叉验证提升模型稳定性和泛化能力.

kf = KFold(n_splits=5, shuffle=True, random_state=42)
test_pred = np.zeros(len(X_test))

for fold, (train_idx, val_idx) in enumerate(kf.split(X)):
    print(f"Training Fold {fold+1} ...")
    
    model = CatBoostClassifier(
        iterations=3000, learning_rate=0.02, depth=7,
        l2_leaf_reg=3, random_strength=1.5, 
        bagging_temperature=0.6, border_count=254,
        loss_function='Logloss', eval_metric='AUC',
        bootstrap_type='Bayesian', od_type='Iter',
        od_wait=80, random_seed=42, task_type='CPU'
    )
    
    model.fit(
        X.iloc[train_idx], y.iloc[train_idx],
        cat_features=cat_cols
    )
    
    test_pred += model.predict_proba(X_test)[:, 1] / 5

print("5-Fold 训练完成！")

2.6 结果生成与提交

生成提交文件

submit = pd.DataFrame({
    'ID': test_processed['ID'],
    'Prediction': test_pred
})
submit.to_csv("submit.csv", index=False)

最高测试得分0.9334422974836348