feat:修改ros回归模型训练任务，添加样本权重

recommend-model-produce/src/main/scala/com/tzld/piaoquan/recommend/model/recsys_01_ros_reg_weight_xgb_train.scala

@@ -0,0 +1,196 @@
+package com.tzld.piaoquan.recommend.model
+
+import com.alibaba.fastjson.JSON
+import com.tzld.piaoquan.recommend.model.produce.util.RosUtil
+import com.tzld.piaoquan.recommend.utils.{MyHdfsUtils, ParamUtils}
+import ml.dmlc.xgboost4j.scala.spark.XGBoostRegressor
+import org.apache.commons.lang.math.NumberUtils
+import org.apache.commons.lang3.StringUtils
+import org.apache.hadoop.io.compress.GzipCodec
+import org.apache.spark.ml.evaluation.RegressionEvaluator
+import org.apache.spark.ml.feature.VectorAssembler
+import org.apache.spark.rdd.RDD
+import org.apache.spark.sql.types.DataTypes
+import org.apache.spark.sql.{Dataset, Row, SparkSession}
+
+import java.util
+import scala.io.Source
+
+/**
+ * 推荐 ros 回归模型，加权损失函数
+ */
+object recsys_01_ros_reg_weight_xgb_train {
+  def main(args: Array[String]): Unit = {
+    val spark = SparkSession
+      .builder()
+      .appName(this.getClass.getName)
+      .getOrCreate()
+    val sc = spark.sparkContext
+
+
+    val param = ParamUtils.parseArgs(args)
+    val featureFile = param.getOrElse("featureFile", "20250306_ros_feature_232.txt")
+    val trainPath = param.getOrElse("trainPath", "/dw/recommend/model/43_recsys_ros_data_bucket/20250301")
+    val testPath = param.getOrElse("testPath", "/dw/recommend/model/43_recsys_ros_data_bucket/20250302")
+    val savePath = param.getOrElse("savePath", "/dw/recommend/model/44_recsys_ros_predict/")
+    val featureFilter = param.getOrElse("featureFilter", "XXXXXX").split(",").filter(_.nonEmpty)
+    val eta = param.getOrElse("eta", "0.01").toDouble
+    val gamma = param.getOrElse("gamma", "0.0").toDouble
+    val max_depth = param.getOrElse("max_depth", "5").toInt
+    val num_round = param.getOrElse("num_round", "100").toInt
+    val num_worker = param.getOrElse("num_worker", "20").toInt
+    val func_object = param.getOrElse("func_object", "reg:squarederror")
+    val func_metric = param.getOrElse("func_metric", "rmse")
+    val repartition = param.getOrElse("repartition", "20").toInt
+    val modelPath = param.getOrElse("modelPath", "/dw/recommend/model/45_recommend_model/20250310_ros_reg_1000")
+    val modelFile = param.getOrElse("modelFile", "model.tar.gz")
+
+    val loader = getClass.getClassLoader
+    val resourceUrl = loader.getResource(featureFile)
+    val content =
+      if (resourceUrl != null) {
+        val content = Source.fromURL(resourceUrl).getLines().mkString("\n")
+        Source.fromURL(resourceUrl).close()
+        content
+      } else {
+        ""
+      }
+    println(content)
+
+    val features = content.split("\n")
+      .map(r => r.replace(" ", "").replaceAll("\n", ""))
+      .filter(r => r.nonEmpty || !featureFilter.contains(r))
+    println("features.size=" + features.length)
+
+    val trainData = createData(
+      sc.textFile(trainPath),
+      features
+    )
+    println("recsys ros:train data size：" + trainData.count())
+
+    var fields = Array(
+      DataTypes.createStructField("label", DataTypes.DoubleType, true)
+    ) ++ features.map(f => DataTypes.createStructField(f, DataTypes.DoubleType, true))
+
+    fields = fields ++ Array(
+      DataTypes.createStructField("logKey", DataTypes.StringType, true),
+      DataTypes.createStructField("weight", DataTypes.DoubleType, true)
+    )
+    val schema = DataTypes.createStructType(fields)
+
+
+    val trainDataSet: Dataset[Row] = spark.createDataFrame(trainData, schema)
+    val vectorAssembler = new VectorAssembler().setInputCols(features).setOutputCol("features")
+    val xgbInput = vectorAssembler.transform(trainDataSet).select("features", "label", "weight").persist()
+    val xgbRegressor = new XGBoostRegressor()
+      .setEta(eta)
+      .setGamma(gamma)
+      .setMissing(0.0f)
+      .setMaxDepth(max_depth)
+      .setNumRound(num_round)
+      .setSubsample(0.8)
+      .setColsampleBytree(0.8)
+      .setObjective(func_object)
+      .setEvalMetric(func_metric)
+      .setFeaturesCol("features")
+      .setLabelCol("label")
+      .setWeightCol("weight")
+      .setNthread(1)
+      .setNumWorkers(num_worker)
+      .setSeed(2024)
+      .setMinChildWeight(1)
+    val model = xgbRegressor.fit(xgbInput)
+
+    if (modelPath.nonEmpty && modelFile.nonEmpty) {
+      model.write.overwrite.save(modelPath)
+    }
+
+    // 评测数据
+    val testData = createData(
+      sc.textFile(testPath),
+      features
+    )
+    val testDataSet = spark.createDataFrame(testData, schema)
+    println("recsys ros testDataSet schema")
+    testDataSet.printSchema()
+
+    val testDataSetTrans = vectorAssembler.transform(testDataSet).select("features", "label", "logKey")
+    println("recsys ros testDataSetTrans schema")
+    testDataSetTrans.printSchema()
+
+
+    val predictions = model.transform(testDataSetTrans)
+
+    // 保存评估结果
+    println("recsys ros:columns:" + predictions.columns.mkString(",")) //[label, features, prediction]
+    val saveData = predictions.select("label", "prediction", "logKey").rdd
+      .map(r => {
+        (r.get(0), r.get(1), r.get(2)).productIterator.mkString("\t")
+      })
+    val hdfsPath = savePath
+    if (hdfsPath.nonEmpty && hdfsPath.startsWith("/dw/recommend/model/")) {
+      println("删除路径并开始数据写入:" + hdfsPath)
+      MyHdfsUtils.delete_hdfs_path(hdfsPath)
+      saveData.repartition(repartition).saveAsTextFile(hdfsPath, classOf[GzipCodec])
+    } else {
+      println("路径不合法，无法写入:" + hdfsPath)
+    }
+
+    // 计算rmse
+    val evaluator = new RegressionEvaluator()
+      .setLabelCol("label")
+      .setPredictionCol("prediction")
+      .setMetricName("rmse")
+    val rmse = evaluator.evaluate(predictions.select("label", "prediction"))
+    println("recsys nor: rmse:" + rmse)
+
+    sc.textFile(hdfsPath).map(r => {
+        val rList = r.split("\t")
+        val vid = JSON.parseObject(rList(2)).getString("vid")
+        val label = rList(0).toDouble
+        val score = rList(1).toDouble
+        (vid, (1, label, RosUtil.inverseLog(label), score, RosUtil.inverseLog(score)))
+      }).reduceByKey {
+        case (c1, c2) => (
+          c1._1 + c2._1,
+          c1._2 + c2._2,
+          c1._3 + c2._3,
+          c1._4 + c2._4,
+          c1._5 + c2._5
+        )
+      }.map {
+        case (vid, (all, label, originLabel, score, originScore)) =>
+          (vid, all, label / all, score / all, originLabel / all, originScore / all)
+      }
+      .collect()
+      .sortBy(-_._2)
+      .take(200)
+      .map(_.productIterator.mkString("\t"))
+      .foreach(println)
+
+  }
+
+  def createData(data: RDD[String], features: Array[String]): RDD[Row] = {
+    data.map(r => {
+      val line: Array[String] = StringUtils.split(r, '\t')
+      val logKey = line(0)
+
+      val label: Double = NumberUtils.toDouble(line(1))
+
+      val map: util.Map[String, Double] = new util.HashMap[String, Double]
+      for (i <- 2 until line.length) {
+        val fv: Array[String] = StringUtils.split(line(i), ':')
+        map.put(fv(0), NumberUtils.toDouble(fv(1), 0.0))
+      }
+
+      val v: Array[Any] = new Array[Any](features.length + 3)
+      v(0) = label
+      for (i <- 0 until features.length) {
+        v(i + 1) = map.getOrDefault(features(i), 0.0d)
+      }
+      v(features.length + 1) = logKey
+      v(features.length + 2) = if (label == 0) 1.5 else 3.0 // 设置权重
+      Row(v: _*)
+    })
+  }
+}