Optimize argtopk

optimizing-100k.json

@@ -12,11 +12,11 @@
         "2.k10u1v864": 0.24232304952129619,
         "3.k10u1v886": 0,
         "4.PredUser1Item1": 4.319093503763853,
-        "5.PredUser327Item2": 2.6994178006921192,
-        "6.Mae": 0.8287277961963542
+        "5.PredUser327Item2": 2.9542350642575563,
+        "6.Mae": 0.8311356402216432
     },
     "BR.2": {
-        "average (ms)": 2748.0986,
+        "average (ms)": 6474.2568,
         "stddev (ms)": 0
     }
 }
\ No newline at end of file

src/main/scala/shared/predictions.scala

@@ -176,17 +176,13 @@ package object predictions
 
   //take k nearest users by similarity excluding itself
   def knn(user : Int, k: Int, similarities : DenseMatrix[Double]) : Array[Int] = {
-    val row = similarities(user, ::)
-    val row_index = row.t.toArray.zipWithIndex
-    val row_sorted = row_index.sortBy(- _._1).take(k+1)
     //first element is itself so take the tail
-    return row_sorted.map(_._2).tail
+    return argtopk(similarities(::,user),k+1).toArray.tail
   }
 
   def predict_knn_similarity(similarities : DenseMatrix[Double], k : Int): DenseMatrix[Double] = {
     for (x <- 0 to similarities.rows - 1){
       val index_user = knn(x, k, similarities)
-
       for(y <- 0 to similarities.cols - 1) {
         // if y is not an index of the largest similarity coefficients it is changed to 0
         if (!index_user.contains(y)) similarities(x,y) = 0.0
@@ -232,8 +228,7 @@ package object predictions
       iterator.map(user=>{
         val ratings = broadcast.value
         val similarities = ratings * ratings.t(::,user)
-        val row_index = similarities.toArray.zipWithIndex
-        val res = row_index.sortBy(- _._1).take(k+1).map(_._2).tail
+        val res=argtopk(similarities,k+1).toArray.tail
         val res_form = res.map(user_2 => (user,user_2,similarities(user_2)))
       res_form})}).collect()