feat:添加新的训练样本生成脚本

src/main/scala/com/aliyun/odps/spark/examples/makedata_recsys/v20250411/makedata_recsys_41_str_train_data_sample_20250411.scala

@@ -0,0 +1,302 @@
+package com.aliyun.odps.spark.examples.makedata_recsys.v20250411
+
+import com.alibaba.fastjson.{JSON, JSONObject}
+import com.aliyun.odps.TableSchema
+import com.aliyun.odps.data.Record
+import com.aliyun.odps.spark.examples.myUtils.{MyDateUtils, MyHdfsUtils, ParamUtils, env}
+import examples.extractor.{ExtractorUtils, RankExtractorFeature_20240530}
+import examples.utils.StatisticsUtil
+import org.apache.hadoop.io.compress.GzipCodec
+import org.apache.spark.sql.SparkSession
+import org.xm.Similarity
+
+import scala.collection.JavaConversions._
+import scala.collection.mutable.ArrayBuffer
+import scala.util.Random
+
+/*
+   20240608 提取特征
+ */
+
+object makedata_recsys_41_str_train_data_sample_20250411 {
+  def main(args: Array[String]): Unit = {
+    val spark = SparkSession
+      .builder()
+      .appName(this.getClass.getName)
+      .getOrCreate()
+    val sc = spark.sparkContext
+
+    // 1 读取参数
+    val param = ParamUtils.parseArgs(args)
+    val tablePart = param.getOrElse("tablePart", "64").toInt
+    val beginStr = param.getOrElse("beginStr", "2025041200")
+    val endStr = param.getOrElse("endStr", "2025041223")
+    val savePath = param.getOrElse("savePath", "/dw/recommend/model/41_recsys_str_train_data_v2/")
+    val project = param.getOrElse("project", "loghubods")
+    val table = param.getOrElse("table", "dwd_recsys_alg_sample_all_20250408")
+    val repartition = param.getOrElse("repartition", "32").toInt
+    val fuSampleRate = param.getOrElse("fuSampleRate", "0.05").toDouble
+    val whatLabel = param.getOrElse("whatLabel", "is_share")
+    val whatApps = param.getOrElse("whatApps", "0,4,2,32,17,18,21,22,24,25,26,27,28,29,3,30,31,33,34,35,36").split(",").filter(r => r.nonEmpty).toList
+
+    // 2 读取odps+表信息
+    val odpsOps = env.getODPS(sc)
+
+    // 3 循环执行数据生产
+    val timeRange = MyDateUtils.getDateHourRange(beginStr, endStr)
+    for (dt_hh <- timeRange) {
+      val dt = dt_hh.substring(0, 8)
+      val hh = dt_hh.substring(8, 10)
+      val partition = s"dt=$dt,hh=$hh"
+      println("开始执行partiton:" + partition)
+      val odpsData = odpsOps.readTable(project = project,
+          table = table,
+          partition = partition,
+          transfer = func,
+          numPartition = tablePart)
+        .filter(record => {
+          val page = record.getString("page")
+          val recommendPageType = record.getString("recommendpagetype")
+          val apptype = record.getString("apptype")
+          whatApps.contains(apptype) && StatisticsUtil.isRecommendScene(page, recommendPageType)
+        })
+        .filter(record => {
+          val label = record.getString(whatLabel)
+          "1".equals(label) || new Random().nextDouble() <= fuSampleRate
+        })
+        .map(record => {
+
+          val featureMap = new JSONObject()
+          val ts = if (record.isNull("ts")) 0 else record.getString("ts").toLong
+
+          // a 视频特征
+          val b1: JSONObject = if (record.isNull("b1_feature")) new JSONObject() else
+            JSON.parseObject(record.getString("b1_feature"))
+          val b2: JSONObject = if (record.isNull("b2_feature")) new JSONObject() else
+            JSON.parseObject(record.getString("b2_feature"))
+          val b3: JSONObject = if (record.isNull("b3_feature")) new JSONObject() else
+            JSON.parseObject(record.getString("b3_feature"))
+          val b6: JSONObject = if (record.isNull("b6_feature")) new JSONObject() else
+            JSON.parseObject(record.getString("b6_feature"))
+          val b7: JSONObject = if (record.isNull("b7_feature")) new JSONObject() else
+            JSON.parseObject(record.getString("b7_feature"))
+
+          val b8: JSONObject = if (record.isNull("b8_feature")) new JSONObject() else
+            JSON.parseObject(record.getString("b8_feature"))
+          val b9: JSONObject = if (record.isNull("b9_feature")) new JSONObject() else
+            JSON.parseObject(record.getString("b9_feature"))
+          val b10: JSONObject = if (record.isNull("b10_feature")) new JSONObject() else
+            JSON.parseObject(record.getString("b10_feature"))
+          val b11: JSONObject = if (record.isNull("b11_feature")) new JSONObject() else
+            JSON.parseObject(record.getString("b11_feature"))
+          val b12: JSONObject = if (record.isNull("b12_feature")) new JSONObject() else
+            JSON.parseObject(record.getString("b12_feature"))
+          val b13: JSONObject = if (record.isNull("b13_feature")) new JSONObject() else
+            JSON.parseObject(record.getString("b13_feature"))
+          val b17: JSONObject = if (record.isNull("b17_feature")) new JSONObject() else
+            JSON.parseObject(record.getString("b17_feature"))
+          val b18: JSONObject = if (record.isNull("b18_feature")) new JSONObject() else
+            JSON.parseObject(record.getString("b18_feature"))
+          val b19: JSONObject = if (record.isNull("b19_feature")) new JSONObject() else
+            JSON.parseObject(record.getString("b19_feature"))
+
+
+          val origin_data = List(
+            (b1, b2, b3, "b123"), (b1, b6, b7, "b167"),
+            (b8, b9, b10, "b8910"), (b11, b12, b13, "b111213"),
+            (b17, b18, b19, "b171819")
+          )
+          for ((b_1, b_2, b_3, prefix1) <- origin_data) {
+            for (prefix2 <- List(
+              "1h", "2h", "3h", "4h", "12h", "1d", "3d", "7d"
+            )) {
+              val exp = if (b_1.isEmpty) 0D else b_1.getIntValue("exp_pv_" + prefix2).toDouble
+              val share = if (b_2.isEmpty) 0D else b_2.getIntValue("share_pv_" + prefix2).toDouble
+              val returns = if (b_3.isEmpty) 0D else b_3.getIntValue("return_uv_" + prefix2).toDouble
+              val f1 = RankExtractorFeature_20240530.calDiv(share, exp)
+              val f2 = RankExtractorFeature_20240530.calLog(share)
+              val f3 = RankExtractorFeature_20240530.calDiv(returns, exp)
+              val f4 = RankExtractorFeature_20240530.calLog(returns)
+              val f5 = f3 * f4
+              val f6 = RankExtractorFeature_20240530.calDiv(returns, share)
+              featureMap.put(prefix1 + "_" + prefix2 + "_" + "STR", f1)
+              featureMap.put(prefix1 + "_" + prefix2 + "_" + "log(share)", f2)
+              featureMap.put(prefix1 + "_" + prefix2 + "_" + "ROV", f3)
+              featureMap.put(prefix1 + "_" + prefix2 + "_" + "log(return)", f4)
+              featureMap.put(prefix1 + "_" + prefix2 + "_" + "ROV*log(return)", f5)
+              featureMap.put(prefix1 + "_" + prefix2 + "_" + "ROS", f6)
+            }
+          }
+
+          val video_info: JSONObject = if (record.isNull("v1_feature")) new JSONObject() else
+            JSON.parseObject(record.getString("v1_feature"))
+          featureMap.put("total_time", if (video_info.containsKey("total_time")) video_info.getIntValue("total_time").toDouble else 0D)
+          featureMap.put("bit_rate", if (video_info.containsKey("bit_rate")) video_info.getIntValue("bit_rate").toDouble else 0D)
+
+          val c1: JSONObject = if (record.isNull("c1_feature")) new JSONObject() else
+            JSON.parseObject(record.getString("c1_feature"))
+          if (c1.nonEmpty) {
+            featureMap.put("playcnt_6h", if (c1.containsKey("playcnt_6h")) c1.getIntValue("playcnt_6h").toDouble else 0D)
+            featureMap.put("playcnt_1d", if (c1.containsKey("playcnt_1d")) c1.getIntValue("playcnt_1d").toDouble else 0D)
+            featureMap.put("playcnt_3d", if (c1.containsKey("playcnt_3d")) c1.getIntValue("playcnt_3d").toDouble else 0D)
+            featureMap.put("playcnt_7d", if (c1.containsKey("playcnt_7d")) c1.getIntValue("playcnt_7d").toDouble else 0D)
+          }
+          val c2: JSONObject = if (record.isNull("c2_feature")) new JSONObject() else
+            JSON.parseObject(record.getString("c2_feature"))
+          if (c2.nonEmpty) {
+            featureMap.put("share_pv_12h", if (c2.containsKey("share_pv_12h")) c2.getIntValue("share_pv_12h").toDouble else 0D)
+            featureMap.put("share_pv_1d", if (c2.containsKey("share_pv_1d")) c2.getIntValue("share_pv_1d").toDouble else 0D)
+            featureMap.put("share_pv_3d", if (c2.containsKey("share_pv_3d")) c2.getIntValue("share_pv_3d").toDouble else 0D)
+            featureMap.put("share_pv_7d", if (c2.containsKey("share_pv_7d")) c2.getIntValue("share_pv_7d").toDouble else 0D)
+            featureMap.put("return_uv_12h", if (c2.containsKey("return_uv_12h")) c2.getIntValue("return_uv_12h").toDouble else 0D)
+            featureMap.put("return_uv_1d", if (c2.containsKey("return_uv_1d")) c2.getIntValue("return_uv_1d").toDouble else 0D)
+            featureMap.put("return_uv_3d", if (c2.containsKey("return_uv_3d")) c2.getIntValue("return_uv_3d").toDouble else 0D)
+            featureMap.put("return_uv_7d", if (c2.containsKey("return_uv_7d")) c2.getIntValue("return_uv_7d").toDouble else 0D)
+          }
+
+          val title = if (video_info.containsKey("title")) video_info.getString("title") else ""
+          if (!title.equals("")) {
+            for (key_feature <- List("c3_feature", "c4_feature", "c5_feature", "c6_feature", "c7_feature")) {
+              val c34567: JSONObject = if (record.isNull(key_feature)) new JSONObject() else
+                JSON.parseObject(record.getString(key_feature))
+              for (key_time <- List("tags_1d", "tags_3d", "tags_7d")) {
+                val tags = if (c34567.containsKey(key_time)) c34567.getString(key_time) else ""
+                if (!tags.equals("")) {
+                  val (f1, f2, f3, f4) = funcC34567ForTags(tags, title)
+                  featureMap.put(key_feature + "_" + key_time + "_matchnum", f1)
+                  featureMap.put(key_feature + "_" + key_time + "_maxscore", f3)
+                  featureMap.put(key_feature + "_" + key_time + "_avgscore", f4)
+                }
+              }
+            }
+          }
+
+          val vid = if (record.isNull("vid")) "" else record.getString("vid")
+          if (!vid.equals("")) {
+            for (key_feature <- List("c8_feature", "c9_feature")) {
+              val c89: JSONObject = if (record.isNull(key_feature)) new JSONObject() else
+                JSON.parseObject(record.getString(key_feature))
+              for (key_action <- List("share", "return")) {
+                val cfListStr = if (c89.containsKey(key_action)) c89.getString(key_action) else ""
+                if (!cfListStr.equals("")) {
+                  val cfMap = cfListStr.split(",").map(r => {
+                    val rList = r.split(":")
+                    (rList(0), (rList(1), rList(2), rList(3)))
+                  }).toMap
+                  if (cfMap.contains(vid)) {
+                    val (score, num, rank) = cfMap(vid)
+                    featureMap.put(key_feature + "_" + key_action + "_score", score.toDouble)
+                    featureMap.put(key_feature + "_" + key_action + "_num", num.toDouble)
+                    featureMap.put(key_feature + "_" + key_action + "_rank", 1.0 / rank.toDouble)
+                  }
+                }
+              }
+            }
+          }
+
+          val d1: JSONObject = if (record.isNull("d1_feature")) new JSONObject() else
+            JSON.parseObject(record.getString("d1_feature"))
+          if (d1.nonEmpty) {
+            featureMap.put("d1_exp", if (d1.containsKey("exp")) d1.getString("exp").toDouble else 0D)
+            featureMap.put("d1_return_n", if (d1.containsKey("return_n")) d1.getString("return_n").toDouble else 0D)
+            featureMap.put("d1_rovn", if (d1.containsKey("rovn")) d1.getString("rovn").toDouble else 0D)
+          }
+
+
+          /*
+
+
+          视频：
+          曝光使用pv 分享使用pv 回流使用uv --> 1h 2h 3h 4h 12h 1d 3d 7d
+          STR log(share) ROV log(return) ROV*log(return)
+          40个特征组合
+          整体、整体曝光对应、推荐非冷启root、推荐冷启root、分省份root
+          200个特征值
+
+          视频：
+          视频时长、比特率
+
+          人：
+          播放次数 --> 6h 1d 3d 7d --> 4个
+          带回来的分享pv 回流uv --> 12h 1d 3d 7d --> 8个
+          人+vid-title:
+          播放点/回流点/分享点/累积分享/累积回流 --> 1d 3d 7d --> 匹配数量 语义最高相似度分 语义平均相似度分 --> 45个
+          人+vid-cf
+          基于分享行为/基于回流行为 -->  “分享cf”+”回流点击cf“ 相似分 相似数量 相似rank的倒数 --> 12个
+
+          头部视频：
+          曝光 回流 ROVn 3个特征
+
+          场景：
+          小时 星期 apptype city province pagesource 机器型号
+           */
+
+
+          //4 处理label信息。
+          val labels = new JSONObject
+          for (labelKey <- List("is_share", "share_cnt"
+            , "is_return_1", "return_1_pv", "return_1_uv", "is_return_noself", "return_1_uv_noself"
+            , "is_return_n", "return_n_pv", "return_n_uv", "is_return_n_noself", "return_n_uv_noself"
+            , "new_exposure_cnt")) {
+            if (!record.isNull(labelKey)) {
+              labels.put(labelKey, record.getString(labelKey))
+            }
+          }
+
+
+          val logs = new JSONObject
+          for (logKey <- List("apptype", "pagesource", "mid", "vid", "subsessionid", "page", "abcode", "recomtraceid", "headvideoid", "hotsencetype", "ts")) {
+            if (!record.isNull(logKey)) {
+              logs.put(logKey, record.getString(logKey))
+            }
+          }
+
+          logs.put("hour", ExtractorUtils.getHourByTimestamp(ts))
+          logs.put("day_of_week", ExtractorUtils.getDayOfWeekByTimestamp(ts))
+
+          //5 处理log key表头。
+          val logKey = logs.toString()
+          val labelKey = labels.toString()
+          val featureKey = featureMap.toString()
+          //6 拼接数据，保存。
+          logKey + "\t" + labelKey + "\t" + featureKey
+
+        })
+
+      // 4 保存数据到hdfs
+      val savePartition = dt + hh
+      val hdfsPath = savePath + "/" + savePartition
+      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)
+        odpsData.coalesce(repartition).saveAsTextFile(hdfsPath, classOf[GzipCodec])
+      } else {
+        println("路径不合法，无法写入:" + hdfsPath)
+      }
+    }
+  }
+
+  def func(record: Record, schema: TableSchema): Record = {
+    record
+  }
+
+  def funcC34567ForTags(tags: String, title: String): Tuple4[Double, String, Double, Double] = {
+    // 匹配数量 匹配词 语义最高相似度分 语义平均相似度分
+    val tagsList = tags.split(",")
+    var d1 = 0.0
+    val d2 = new ArrayBuffer[String]()
+    var d3 = 0.0
+    var d4 = 0.0
+    for (tag <- tagsList) {
+      if (title.contains(tag)) {
+        d1 = d1 + 1.0
+        d2.add(tag)
+      }
+      val score = Similarity.conceptSimilarity(tag, title)
+      d3 = if (score > d3) score else d3
+      d4 = d4 + score
+    }
+    d4 = if (tagsList.nonEmpty) d4 / tagsList.size else d4
+    (d1, d2.mkString(","), d3, d4)
+  }
+}