185 lines
6.9 KiB
Python
185 lines
6.9 KiB
Python
from pydantic import BaseModel, Field
|
||
import joblib
|
||
import json
|
||
import pandas as pd
|
||
import numpy as np
|
||
import logging
|
||
import os
|
||
import random
|
||
|
||
# 配置日志
|
||
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
|
||
logger = logging.getLogger(__name__)
|
||
|
||
# ==========================================
|
||
# 1. 定义 Pydantic 数据模型
|
||
# ==========================================
|
||
|
||
class CustomerFeatures(BaseModel):
|
||
"""客户特征输入模型"""
|
||
age: int = Field(ge=18, le=120, description="客户年龄")
|
||
job: str = Field(description="职业")
|
||
marital: str = Field(description="婚姻状况")
|
||
education: str = Field(description="教育程度")
|
||
default: str = Field(pattern="^(yes|no)$", description="是否有违约记录")
|
||
balance: int = Field(description="账户余额")
|
||
housing: str = Field(pattern="^(yes|no)$", description="是否有住房贷款")
|
||
loan: str = Field(pattern="^(yes|no)$", description="是否有个人贷款")
|
||
contact: str = Field(description="联系方式")
|
||
day: int = Field(ge=1, le=31, description="最后联系日")
|
||
month: str = Field(description="最后联系月份")
|
||
campaign: int = Field(ge=1, description="本次活动联系次数")
|
||
pdays: int = Field(description="距离上次联系天数 (-1代表未联系)")
|
||
previous: int = Field(ge=0, description="活动前联系次数")
|
||
poutcome: str = Field(description="上次活动结果")
|
||
|
||
# 注意：我们不包含 duration，因为它是事后变量
|
||
|
||
class Decision(BaseModel):
|
||
"""Agent 输出的结构化决策"""
|
||
risk_score: float = Field(ge=0, le=1, description="预测购买概率 (0-1)")
|
||
customer_segment: str = Field(description="客户分群 (如: 高价值/潜在/沉睡)")
|
||
decision: str = Field(description="建议策略 (如: 立即致电/邮件触达/放弃)")
|
||
actions: list[str] = Field(description="可执行动作清单")
|
||
rationale: str = Field(description="决策依据 (结合模型预测与业务规则)")
|
||
|
||
# ==========================================
|
||
# 2. 定义 Agent 类
|
||
# ==========================================
|
||
|
||
class MarketingAgent:
|
||
def __init__(self, model_path="models/model_artifacts.pkl"):
|
||
self.model_path = model_path
|
||
self.artifacts = None
|
||
self._load_model()
|
||
|
||
def _load_model(self):
|
||
if os.path.exists(self.model_path):
|
||
self.artifacts = joblib.load(self.model_path)
|
||
logger.info(f"Agent 已加载模型: {self.model_path}")
|
||
else:
|
||
logger.warning(f"模型文件不存在: {self.model_path}，Agent 将无法进行预测")
|
||
|
||
def predict_risk(self, features: CustomerFeatures) -> dict:
|
||
"""
|
||
Tool 1: 调用 ML 模型预测购买概率
|
||
"""
|
||
if not self.artifacts:
|
||
return {"score": 0.0, "reason": "Model not loaded"}
|
||
|
||
# 转换输入为 DataFrame
|
||
data = features.model_dump()
|
||
df = pd.DataFrame([data])
|
||
|
||
# 预处理 (使用训练时保存的 encoder)
|
||
# 注意：这里需要严格复现训练时的预处理逻辑
|
||
# 训练时我们做了 Label Encoding
|
||
for col, le in self.artifacts['encoders'].items():
|
||
if col in df.columns:
|
||
# 处理未知类别
|
||
try:
|
||
df[col] = le.transform(df[col].astype(str))
|
||
except:
|
||
# 遇到未知类别，这里简单处理为 0 (或者 mode)
|
||
logger.warning(f"Unknown category in {col}")
|
||
df[col] = 0
|
||
|
||
# 确保列顺序一致
|
||
# 我们训练时用了 X (df.drop(target))
|
||
# 这里需要筛选出 numeric_cols + categorical_cols
|
||
# 简单起见，我们假设 feature names 保存了顺序
|
||
feature_names = self.artifacts['features']
|
||
|
||
# 补齐可能缺失的列
|
||
for col in feature_names:
|
||
if col not in df.columns:
|
||
df[col] = 0
|
||
|
||
X_input = df[feature_names]
|
||
|
||
# 预测
|
||
model = self.artifacts['lgb_model'] # 优先使用 LightGBM
|
||
prob = model.predict_proba(X_input)[0][1]
|
||
|
||
return {
|
||
"score": float(prob),
|
||
"top_features": ["balance", "poutcome"] # 这里简化，实际可用 SHAP
|
||
}
|
||
|
||
def get_strategy(self, score: float) -> dict:
|
||
"""
|
||
Tool 2: 规则引擎/检索工具
|
||
"""
|
||
if score > 0.6:
|
||
return {
|
||
"segment": "高意向 VIP",
|
||
"action_type": "人工介入",
|
||
"templates": ["尊贵的客户，鉴于您...", "专属理财经理一对一服务"]
|
||
}
|
||
elif score > 0.3:
|
||
return {
|
||
"segment": "潜在客户",
|
||
"action_type": "自动化营销",
|
||
"templates": ["你好，近期理财活动...", "点击领取加息券"]
|
||
}
|
||
else:
|
||
return {
|
||
"segment": "低意向群体",
|
||
"action_type": "静默/邮件",
|
||
"templates": ["月度财经摘要"]
|
||
}
|
||
|
||
def run(self, features: CustomerFeatures) -> Decision:
|
||
"""
|
||
Agent 主流程
|
||
"""
|
||
logger.info(f"Agent 正在处理客户: {features.job}, {features.age}岁")
|
||
|
||
# 1. 感知 (调用 ML 工具)
|
||
pred_result = self.predict_risk(features)
|
||
score = pred_result["score"]
|
||
|
||
# 2. 规划 (调用 策略工具)
|
||
strategy = self.get_strategy(score)
|
||
|
||
# 3. 决策 (模拟 LLM 整合)
|
||
# 在真实场景中，这里构建 Prompt 发送给 DeepSeek
|
||
# 这里我们用 Python 逻辑模拟 LLM 的结构化输出能力
|
||
|
||
decision = Decision(
|
||
risk_score=round(score, 4),
|
||
customer_segment=strategy["segment"],
|
||
decision=f"建议采取 {strategy['action_type']}",
|
||
actions=[f"使用话术: {t}" for t in strategy["templates"]],
|
||
rationale=f"模型预测概率为 {score:.1%}，属于{strategy['segment']}。该群体对{strategy['action_type']}转化率较高。"
|
||
)
|
||
|
||
return decision
|
||
|
||
if __name__ == "__main__":
|
||
# 测试 Agent
|
||
agent = MarketingAgent()
|
||
|
||
# 构造一个测试用例
|
||
test_customer = CustomerFeatures(
|
||
age=35,
|
||
job="management",
|
||
marital="married",
|
||
education="tertiary",
|
||
default="no",
|
||
balance=2000,
|
||
housing="yes",
|
||
loan="no",
|
||
contact="cellular",
|
||
day=15,
|
||
month="may",
|
||
campaign=1,
|
||
pdays=-1,
|
||
previous=0,
|
||
poutcome="unknown"
|
||
)
|
||
|
||
result = agent.run(test_customer)
|
||
print("\n=== Agent Decision ===")
|
||
print(result.model_dump_json(indent=2))
|