tweak default parameters

build.sbt

@@ -2,7 +2,7 @@ organization := "megadata"
 
 name := "SpectralLDA-Tensor"
 
-version := "1.3.0.rc12"
+version := "1.3.0.rc13"
 
 scalaVersion := "2.11.12"
 

src/main/scala/megadata/spectralLDA/algorithm/RandNLA.scala

@@ -40,7 +40,7 @@ object RandNLA {
               numDocs: Long,
               firstOrderMoments: DenseVector[Double],
               documents: RDD[(Long, Double, SparseVector[Double])],
-              nIter: Int = 1)
+              nIter: Int)
             (implicit randBasis: RandBasis = Rand)
   : (DenseMatrix[Double], DenseVector[Double]) = {
     assert(vocabSize >= dimK)
@@ -55,7 +55,7 @@ object RandNLA {
     // Universality laws for randomized dimension reduction
     // with applications, S. Oymak and J. A. Tropp. Inform. Inference, Nov. 2017
     // Theorem II on Restricted Minimum Singular Value
-    val projectedDim = math.pow(dimK, 1.1).toInt
+    val projectedDim = math.pow(dimK, 1.15).toInt
 
     // Cache some data
     val extDocs = documents

src/main/scala/megadata/spectralLDA/algorithm/TensorLDA.scala

@@ -27,6 +27,9 @@ import org.apache.log4j.Logger
   *                                   true by default
   * @param numIterationsKrylovMethod  iterations of the Krylov Method for
   *                                   randomised SVD, 2 by default
+  * @param slackDimK                  extra number of orthogonal topic
+  *                                   fragments to discover in order to
+  *                                   compose k topics
   * @param postProcessing             post-processing topic-word distribution
   *                                   matrix by projection into l1-simplex,
   *                                   false by default
@@ -36,7 +39,7 @@ class TensorLDA(dimK: Int,
                 maxIterations: Int = 500,
                 tol: Double = 1e-6,
                 randomisedSVD: Boolean = true,
-                numIterationsKrylovMethod: Int = 2,
+                numIterationsKrylovMethod: Int = 1,
                 slackDimK: Option[Int] = None,
                 postProcessing: Boolean = false) extends Serializable {
   @transient private lazy val logger = Logger.getLogger("TensorLDA")
@@ -48,7 +51,7 @@ class TensorLDA(dimK: Int,
   // As one final discovered topic could be the combination of multiple
   // eigenvectors of M2, we allow redundancy in the number of eigenvectors
   // we compute for M2.
-  val slackK = slackDimK.getOrElse(dimK / 2)
+  val slackK = slackDimK.getOrElse(dimK)
   logger.info(s"Slack of random projection dimension: $slackK")
 
   assert(alpha0 > 0, "The topic concentration alpha0 must be positive.")

src/main/scala/megadata/spectralLDA/datamoments/DataCumulant.scala

@@ -46,7 +46,7 @@ object DataCumulant {
                       alpha0: Double,
                       documents: RDD[(Long, SparseVector[Double])],
                       randomisedSVD: Boolean = true,
-                      numIterationsKrylovMethod: Int = 1)
+                      numIterationsKrylovMethod: Int)
                      (implicit randBasis: RandBasis = Rand)
         : DataCumulant = {
     assert(dimK > 0, "The number of topics dimK must be positive.")

src/test/scala/megadata/spectralLDA/datamoments/DataCumulantTest.scala

@@ -57,7 +57,8 @@ class DataCumulantTest extends FlatSpec with Matchers {
       dimK = 3,
       alpha0 = sum(alpha),
       documentsRDD,
-      randomisedSVD = false
+      randomisedSVD = false,
+      numIterationsKrylovMethod = 0
     )
 
     val expectedM3: DenseMatrix[Double] = expected_whitened_M3(