diff --git a/mllib/src/main/scala/org/apache/spark/ml/recommendation/ALS.scala b/mllib/src/main/scala/org/apache/spark/ml/recommendation/ALS.scala
index 4a5743650c547..60dd7367053e2 100644
--- a/mllib/src/main/scala/org/apache/spark/ml/recommendation/ALS.scala
+++ b/mllib/src/main/scala/org/apache/spark/ml/recommendation/ALS.scala
@@ -379,20 +379,14 @@ class ALSModel private[ml] (
     // We'll force the IDs to be Int. Unfortunately this converts IDs to Int in the output.
     val topKAggregator = new TopByKeyAggregator[Int, Int, Float](num, Ordering.by(_._2))
     val recs = ratings.as[(Int, Int, Float)].groupByKey(_._1).agg(topKAggregator.toColumn)
-      .toDF(srcOutputColumn, "recommendations")
-
-    // There is some performance hit from converting the (Int, Float) tuples to
-    // (dstOutputColumn: Int, rating: Float) structs using .rdd. Need SPARK-16483 for a fix.
-    val schema = new StructType()
-      .add(srcOutputColumn, IntegerType)
-      .add("recommendations",
-        ArrayType(
-          StructType(
-            StructField(dstOutputColumn, IntegerType, nullable = false) ::
-            StructField("rating", FloatType, nullable = false) ::
-            Nil
-      )))
-    recs.sparkSession.createDataFrame(recs.rdd, schema)
+      .toDF("id", "recommendations")
+
+    val arrayType = ArrayType(
+      new StructType()
+        .add(dstOutputColumn, IntegerType)
+        .add("rating", FloatType)
+    )
+    recs.select($"id" as srcOutputColumn, $"recommendations" cast arrayType)
   }
 }