predictive-intelligence

ServiceNow 预测智能

预测智能使用机器学习自动执行分类、路由和推荐。

PI 功能

功能	用例
分类	自动对事件和案例分类
相似度	查找相似记录
聚类	对相关项目分组
回归	预测数值
推荐	推荐后续操作

关键表

表	用途
ml_solution	ML 解决方案定义
ml_solution_definition	解决方案配置
ml_capability_definition	功能设置
ml_model	已训练的模型
ml_prediction_result	预测结果

分类（ES5）

配置分类解决方案

// Create classification solution (ES5 ONLY!)
// Note: Usually done via UI, shown for understanding

var solution = new GlideRecord("ml_solution")
solution.initialize()

solution.setValue("name", "Incident Category Classifier")
solution.setValue("label", "Incident Category Classifier")
solution.setValue("table", "incident")
solution.setValue("active", true)

// Capability type
solution.setValue("capability", "classification")

// Target field to predict
solution.setValue("target_field", "category")

// Input fields for training
solution.setValue("input_fields", "short_description,description")

solution.insert()

获取分类预测

// Get classification prediction for record (ES5 ONLY!)
function getClassificationPrediction(tableName, recordSysId, solutionName) {
  var predictor = new sn_ml.ClassificationPredictor(solutionName)

  var gr = new GlideRecord(tableName)
  if (!gr.get(recordSysId)) {
    return null
  }

  try {
    var result = predictor.predict(gr)

    return {
      predicted_value: result.getPredictedValue(),
      confidence: result.getConfidence(),
      top_predictions: result.getTopPredictions(5),
    }
  } catch (e) {
    gs.error("Prediction failed: " + e.message)
    return null
  }
}

将预测应用于记录

// Auto-apply classification prediction (ES5 ONLY!)
// Business Rule: before, insert, incident

;(function executeRule(current, previous) {
  // Skip if already categorized
  if (current.category) {
    return
  }

  var solutionName = "incident_category_classifier"

  try {
    var predictor = new sn_ml.ClassificationPredictor(solutionName)
    var result = predictor.predict(current)

    // Only apply if confidence is high enough
    if (result.getConfidence() >= 0.8) {
      current.category = result.getPredictedValue()
      current.work_notes =
        "Category auto-assigned by Predictive Intelligence " +
        "(Confidence: " +
        Math.round(result.getConfidence() * 100) +
        "%)"
    }
  } catch (e) {
    gs.warn("Classification prediction failed: " + e.message)
  }
})(current, previous)

相似度（ES5）

查找相似记录

// Find similar incidents (ES5 ONLY!)
function findSimilarIncidents(incidentSysId, maxResults) {
  maxResults = maxResults || 5

  var incident = new GlideRecord("incident")
  if (!incident.get(incidentSysId)) {
    return []
  }

  try {
    var similarity = new sn_ml.SimilarityPredictor("incident_similarity")
    var results = similarity.findSimilar(incident, maxResults)

    var similar = []
    for (var i = 0; i < results.length; i++) {
      var match = results[i]
      similar.push({
        sys_id: match.getRecordSysId(),
        similarity_score: match.getSimilarityScore(),
        record: match.getRecord(),
      })
    }

    return similar
  } catch (e) {
    gs.error("Similarity search failed: " + e.message)
    return []
  }
}

相似记录小组件

// Widget Server Script for similar records (ES5 ONLY!)
;(function () {
  if (!input || !input.table || !input.sys_id) {
    data.similar = []
    return
  }

  var solutionName = input.table + "_similarity"

  try {
    var gr = new GlideRecord(input.table)
    if (!gr.get(input.sys_id)) {
      data.similar = []
      return
    }

    var similarity = new sn_ml.SimilarityPredictor(solutionName)
    var results = similarity.findSimilar(gr, 5)

    data.similar = []
    for (var i = 0; i < results.length; i++) {
      var match = results[i]
      var record = match.getRecord()

      data.similar.push({
        sys_id: match.getRecordSysId(),
        score: Math.round(match.getSimilarityScore() * 100),
        number: record.getValue("number"),
        short_description: record.getValue("short_description"),
        state: record.state.getDisplayValue(),
      })
    }
  } catch (e) {
    data.error = "Similarity search unavailable"
    data.similar = []
  }
})()

聚类（ES5）

获取聚类分配

// Get cluster for record (ES5 ONLY!)
function getClusterAssignment(tableName, recordSysId, solutionName) {
  var gr = new GlideRecord(tableName)
  if (!gr.get(recordSysId)) {
    return null
  }

  try {
    var clustering = new sn_ml.ClusteringPredictor(solutionName)
    var result = clustering.predict(gr)

    return {
      cluster_id: result.getClusterId(),
      cluster_label: result.getClusterLabel(),
      confidence: result.getConfidence(),
    }
  } catch (e) {
    gs.error("Clustering failed: " + e.message)
    return null
  }
}

分析聚类

// Get cluster statistics (ES5 ONLY!)
function getClusterStats(solutionName) {
  var stats = []

  var cluster = new GlideRecord("ml_cluster")
  cluster.addQuery("solution.name", solutionName)
  cluster.query()

  while (cluster.next()) {
    stats.push({
      cluster_id: cluster.getValue("cluster_id"),
      label: cluster.getValue("label"),
      size: parseInt(cluster.getValue("record_count"), 10),
      keywords: cluster.getValue("keywords"),
    })
  }

  return stats
}

训练模型（ES5）

触发模型训练

// Trigger retraining of ML solution (ES5 ONLY!)
function retrainSolution(solutionName) {
  var solution = new GlideRecord("ml_solution")
  if (!solution.get("name", solutionName)) {
    gs.error("Solution not found: " + solutionName)
    return false
  }

  try {
    // Queue training job
    var trainer = new sn_ml.MLTrainer()
    trainer.train(solution.getUniqueValue())

    gs.info("Training queued for solution: " + solutionName)
    return true
  } catch (e) {
    gs.error("Training failed: " + e.message)
    return false
  }
}

检查训练状态

// Check model training status (ES5 ONLY!)
function getTrainingStatus(solutionName) {
  var model = new GlideRecord("ml_model")
  model.addQuery("solution.name", solutionName)
  model.orderByDesc("sys_created_on")
  model.setLimit(1)
  model.query()

  if (model.next()) {
    return {
      model_id: model.getUniqueValue(),
      status: model.getValue("state"),
      accuracy: model.getValue("accuracy"),
      trained_on: model.getValue("sys_created_on"),
      record_count: model.getValue("training_record_count"),
    }
  }

  return null
}

预测结果（ES5）

存储预测反馈

// Record prediction feedback for model improvement (ES5 ONLY!)
function recordPredictionFeedback(predictionSysId, wasCorrect, actualValue) {
  var prediction = new GlideRecord("ml_prediction_result")
  if (!prediction.get(predictionSysId)) {
    return false
  }

  prediction.setValue("feedback", wasCorrect ? "correct" : "incorrect")
  prediction.setValue("actual_value", actualValue)
  prediction.setValue("feedback_date", new GlideDateTime())
  prediction.setValue("feedback_user", gs.getUserID())

  prediction.update()

  return true
}

分析预测准确性

// Get prediction accuracy stats (ES5 ONLY!)
function getPredictionAccuracy(solutionName, days) {
  days = days || 30

  var startDate = new GlideDateTime()
  startDate.addDaysLocalTime(-days)

  var ga = new GlideAggregate("ml_prediction_result")
  ga.addQuery("solution.name", solutionName)
  ga.addQuery("sys_created_on", ">=", startDate)
  ga.addNotNullQuery("feedback")
  ga.addAggregate("COUNT")
  ga.groupBy("feedback")
  ga.query()

  var stats = { correct: 0, incorrect: 0 }

  while (ga.next()) {
    var feedback = ga.getValue("feedback")
    var count = parseInt(ga.getAggregate("COUNT"), 10)
    stats[feedback] = count
  }

  var total = stats.correct + stats.incorrect
  stats.accuracy = total > 0 ? Math.round((stats.correct / total) * 100) : 0
  stats.total = total

  return stats
}

MCP 工具集成

可用工具

工具	用途
snow_query_table	查询 ML 表
snow_execute_script	测试预测
snow_ml_predict	使用已训练的模型运行预测
snow_list_pi_solutions	列出预测智能解决方案
snow_train_pi_solution	训练 / 重新训练解决方案

示例工作流

// 1. Query ML solutions
await snow_query_table({
  table: "ml_solution",
  query: "active=true",
  fields: "name,table,capability,target_field",
})

// 2. Check model status
await snow_query_table({
  table: "ml_model",
  query: "solution.active=true",
  fields: "solution,state,accuracy,sys_created_on",
})

// 3. Test prediction
await snow_execute_script({
  script: `
        var result = getClassificationPrediction('incident', 'inc_sys_id', 'incident_classifier');
        gs.info(JSON.stringify(result));
    `,
})

最佳实践

1. 高质量数据 - 清理训练数据至关重要
2. 特征选择 - 选择相关的输入字段
3. 置信度阈值 - 仅应用高置信度预测
4. 反馈循环 - 收集用户反馈
5. 定期重新训练 - 定期更新模型
6. 监控准确性 - 跟踪预测性能
7. 回退方案 - 预测失败时应有人工处理流程
8. 仅限 ES5 - 不使用现代 JavaScript 语法