Elki GDBSCAN Java/Scala - how to modify the CorePredicate

Question

How is the generalised dbscan (gdbscan) in elki implemented in Java/Scala? I am currently trying to find an efficient way to implement a weighted dbscan on elki to offset the inefficiencies coming from the sklearn implementation of the weighted dbscan.

The reason I am doing this at the moment is because the sklearn simply sucks for implementing the dbscan on clusters on datasets on the terabyte scale (on the cloud, which in this case I am).

For example, I have made the following code with the database creation function and the dbscan function that reads an array of arrays, and spits out the indices of the cluster indices.

/* Libraries imported from the ELKI library - https://elki-project.github.io/releases/current/doc/overview-summary.html */
import de.lmu.ifi.dbs.elki.algorithm.clustering.kmeans.KMeansElkan 
import de.lmu.ifi.dbs.elki.data.model.{ClusterModel, DimensionModel, KMeansModel, Model} 
import de.lmu.ifi.dbs.elki.data.model
import de.lmu.ifi.dbs.elki.data.{Clustering, DoubleVector, NumberVector}
import de.lmu.ifi.dbs.elki.database.{Database, StaticArrayDatabase}
import de.lmu.ifi.dbs.elki.datasource.ArrayAdapterDatabaseConnection
import de.lmu.ifi.dbs.elki.distance.distancefunction.minkowski.SquaredEuclideanDistanceFunction
import de.lmu.ifi.dbs.elki.distance.distancefunction.minkowski.EuclideanDistanceFunction
import de.lmu.ifi.dbs.elki.distance.distancefunction.NumberVectorDistanceFunction
import de.lmu.ifi.dbs.elki.algorithm.clustering.DBSCAN

// Imports for generalized DBSCAN
import de.lmu.ifi.dbs.elki.algorithm.clustering.gdbscan // Generalized dbscan function here required for weighted dbscan
import de.lmu.ifi.dbs.elki.algorithm.clustering.gdbscan.CorePredicate // THIS IS IMPORTANT TO GET GENERALIZED DBSCAN
import de.lmu.ifi.dbs.elki.algorithm.clustering.gdbscan.GeneralizedDBSCAN
import de.lmu.ifi.dbs.elki.utilities.ELKIBuilder

import de.lmu.ifi.dbs.elki.database.relation.Relation
import de.lmu.ifi.dbs.elki.datasource.DatabaseConnection
import de.lmu.ifi.dbs.elki.database.ids.DBIDIter
import de.lmu.ifi.dbs.elki.index.tree.metrical.covertree.SimplifiedCoverTree
import de.lmu.ifi.dbs.elki.data.{`type`=>TYPE} // Need to import in this way as 'type' is a class method in Scala
import de.lmu.ifi.dbs.elki.index.tree.spatial.rstarvariants.rstar.RStarTreeFactory // Important

def createDatabaseWeighted(data: Array[Array[Double]], distanceFunction: NumberVectorDistanceFunction[NumberVector]): Database = {
  val indexFactory = new SimplifiedCoverTree.Factory[NumberVector](distanceFunction, 0, 30)
  // Create a database
  val db = new StaticArrayDatabase(new ArrayAdapterDatabaseConnection(data), java.util.Arrays.asList(indexFactory))
  // Load the data into the database
  val CustomPredicate = CorePredicate
  db
}

def dbscanClusteringOriginalTest(data: Array[Array[Double]], distanceFunction: NumberVectorDistanceFunction[NumberVector] = SquaredEuclideanDistanceFunction.STATIC, epsilon: Double = 10, minpts: Int = 10) = {
  // Use the same `distanceFunction` for the database and DBSCAN <- is it required??
  val db = createDatabaseWeighted(data, distanceFunction)
  val rel = db.getRelation(TYPE.TypeUtil.NUMBER_VECTOR_FIELD) // Create the required relational database
  val dbscan = new DBSCAN[DoubleVector](distanceFunction, epsilon, minpts) // Epsilon and minpoints needed - either you define in the function input, or will use default values
  val result: Clustering[Model] = dbscan.run(db)
  var ClusterCounter = 0 // Indexing the number of datapoints allocated from DBSCAN

  result.getAllClusters.asScala.zipWithIndex.foreach { case (cluster, idx) =>
    println("The type is " + cluster.getNameAutomatic)
    /* Isolate only the clusters and store the median from the DBSCAN results */
    if (cluster.getNameAutomatic == "Cluster" || cluster.getNameAutomatic == "Noise") {
      ClusterCounter += 1
      val ArrayMedian =  Array[Double]()
      println(s"# $idx: ${cluster.getNameAutomatic}")
      println(s"Size: ${cluster.size()}")
      println(s"Model: ${cluster.getModel}")
      println(s"ids: ${cluster.getIDs.iter().toString}")
    }
  }
}

I can get this to run quite efficiently, but I am currently struggling on how I can get a similar effect with the gdbscan function. For example, there was an answer that suggested that this could be done by modifying the CorePredicate on ELKI (sample_weight option in the ELKI implementation of DBSCAN) but I am not sure how this could be implemented.

Any pointers would be highly appreciated!

Answer 1

Implement your own GDBSCAN core predicate.

Rather than counting neighbors as in the standard implementation, add their weights.

Then you have weighted DBSCAN.