info.debatty.java.graphs.build

Class NNDescent<T>

java.lang.Object

info.debatty.java.graphs.build.GraphBuilder<T>

info.debatty.java.graphs.build.NNDescent<T>

Type Parameters:

T - The type of nodes

All Implemented Interfaces:

Serializable

public class NNDescent<T>
extends GraphBuilder<T>

Implementation of NN-Descent k-nn graph building algorithm. Based on the paper "Efficient K-Nearest Neighbor Graph Construction for Generic Similarity Measures" by Dong et al. http://www.cs.princeton.edu/cass/papers/www11.pdf NN-Descent works by iteratively exploring the neighbors of neighbors... It is not suitable for small datasets (less than 500 items)!

Author:

Thibault Debatty

See Also:

Serialized Form

Field Summary

Fields inherited from class info.debatty.java.graphs.build.GraphBuilder

DEFAULT_K

Constructor Summary

Constructors

Constructor and Description
NNDescent()

Method Summary

Modifier and Type	Method and Description
protected Graph<T>	computeGraph(List<T> nodes, int k, SimilarityInterface<T> similarity)
double	getDelta()
int	getMaxIterations()
double	getRho()
protected Set<Edge>	getSetInstance(int size) Return the correct instance of Set.
protected Graph<T>	nndescent(List<T> nodes, SimilarityInterface<T> similarity)
protected ArrayList<T>	pickFalses(T node, NeighborList neighbor_list)
protected ArrayList<T>	pickTruesAndMark(T node, NeighborList neighbor_list) Pick new neighbors with a probability of rho, and mark them as false.
protected NeighborList	randomNeighborList(List<T> nodes, T for_node)
protected HashMap<T,ArrayList<T>>	reverse(List<T> nodes, Map<T,ArrayList<T>> lists) Reverse NN array R[v] is the list of elements (u) for which v is a neighbor (v is in B[u]).
protected ArrayList<T>	sample(ArrayList<T> nodes, int count)
void	setDelta(double delta) Early termination coefficient.
void	setMaxIterations(int max_iterations) Set the maximum number of iterations.
void	setRho(double rho) Sampling coefficient.
protected ArrayList<T>	union(ArrayList<T> l1, ArrayList<T> l2)
protected int	updateNL(NeighborList nl, T n, double similarity)

Methods inherited from class info.debatty.java.graphs.build.GraphBuilder

computeGraph, getComputedSimilarities, getK, getSimilarity, readFile, setK, setSimilarity, test

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

NNDescent

public NNDescent()

Method Detail

nndescent

protected final Graph<T> nndescent(List<T> nodes,
                                   SimilarityInterface<T> similarity)

getSetInstance

protected final Set<Edge> getSetInstance(int size)

Return the correct instance of Set. Must be synchronized for multi-threaded implementation of NNDescent.

Returns:

getRho

public final double getRho()

Returns:

setRho

public final void setRho(double rho)

Sampling coefficient. In interval ]0, 1.0] Typical value for fast computation is 0.5 Use 1.0 for precise computation Default is 0.5

Parameters:

rho -

getDelta

public final double getDelta()

Returns:

setDelta

public final void setDelta(double delta)

Early termination coefficient. The algorithm stops when less than this proportion of edges are modified Should be in ]0, 1.0[ Default is 0.001

Parameters:

delta -

getMaxIterations

public final int getMaxIterations()

Returns:

setMaxIterations

public final void setMaxIterations(int max_iterations)

Set the maximum number of iterations. Default is Integer.MAX_VALUE

Parameters:

max_iterations -

union

protected final ArrayList<T> union(ArrayList<T> l1,
                                   ArrayList<T> l2)

randomNeighborList

protected final NeighborList randomNeighborList(List<T> nodes,
                                                T for_node)

pickFalses

protected final ArrayList<T> pickFalses(T node,
                                        NeighborList neighbor_list)

pickTruesAndMark

protected final ArrayList<T> pickTruesAndMark(T node,
                                              NeighborList neighbor_list)

Pick new neighbors with a probability of rho, and mark them as false.

Parameters:

node -

neighbor_list -

Returns:

reverse

protected final HashMap<T,ArrayList<T>> reverse(List<T> nodes,
                                                Map<T,ArrayList<T>> lists)

Reverse NN array R[v] is the list of elements (u) for which v is a neighbor (v is in B[u]).

Parameters:

nodes -

lists -

Returns:

sample

protected final ArrayList<T> sample(ArrayList<T> nodes,
                                    int count)

updateNL

protected final int updateNL(NeighborList nl,
                             T n,
                             double similarity)

Parameters:

nl -

n -

similarity -

Returns:

computeGraph

protected Graph<T> computeGraph(List<T> nodes,
                                int k,
                                SimilarityInterface<T> similarity)

Specified by:

computeGraph in class GraphBuilder<T>