ML Models API

机器学习模型管理 API，用于模型训练、部署和预测。

端点列表

方法	端点	说明
POST	/api/ml-models	创建模型
GET	/api/ml-models	获取模型列表
GET	/api/ml-models/{id}	获取模型详情
PUT	/api/ml-models/{id}	更新模型
DELETE	/api/ml-models/{id}	删除模型
POST	/api/ml-models/{id}/train	启动训练
POST	/api/ml-models/{id}/deploy	部署模型
POST	/api/ml-models/{id}/predict	预测请求

模型类型

类型	说明	应用场景
CLASSIFICATION	分类模型	流失预测、作弊检测
REGRESSION	回归模型	LTV 预测、风险评分
CLUSTERING	聚类模型	玩家分群
ANOMALY_DETECTION	异常检测	异常行为识别
TIME_SERIES	时间序列	趋势预测

模型状态

状态	说明
DRAFT	草稿
TRAINING	训练中
EVALUATING	评估中
DEPLOYED	已部署
ARCHIVED	已归档

创建模型

POST /api/ml-models
Content-Type: application/json
Authorization: Bearer {token}

{
  "gameId": "game_abc123",
  "name": "Churn Prediction Model",
  "type": "CLASSIFICATION",
  "algorithm": "Random Forest",
  "framework": "scikit-learn",
  "description": "预测玩家流失概率"
}

响应:

{
  "id": "ml_abc123",
  "gameId": "game_abc123",
  "name": "Churn Prediction Model",
  "type": "CLASSIFICATION",
  "status": "DRAFT",
  "version": 1,
  "createdAt": "2024-01-01T00:00:00Z"
}

启动训练

POST /api/ml-models/{id}/train
Content-Type: application/json
Authorization: Bearer {token}

{
  "datasetConfig": {
    "source": "clickhouse",
    "query": "SELECT * FROM events WHERE event_date > '2024-01-01'",
    "features": ["play_time", "level", "purchases"],
    "label": "churned"
  },
  "hyperparameters": {
    "n_estimators": 100,
    "max_depth": 10
  }
}

部署模型

POST /api/ml-models/{id}/deploy
Content-Type: application/json
Authorization: Bearer {token}

{
  "servingEndpoint": "http://ml-service:8080/predict",
  "canaryDeployment": false
}

预测请求

POST /api/ml-models/{id}/predict
Content-Type: application/json
Authorization: Bearer {token}

{
  "features": {
    "play_time": 120,
    "level": 15,
    "purchases": 3
  }
}

响应:

{
  "prediction": "churn",
  "probability": 0.85,
  "confidence": 0.92,
  "latencyMs": 15
}

模型监控

• 预测量 - 每日预测请求数
• 延迟 - P50/P95/P99 延迟
• 准确率 - 基于反馈的准确率
• 漂移检测 - 模型性能下降告警

A/B 测试

支持模型 A/B 测试：

POST /api/ml-models/{id}/ab-test
Content-Type: application/json
Authorization: Bearer {token}

{
  "baselineModelId": "ml_baseline",
  "trafficSplit": 50
}