nearest neighbour search kdTree

Question

To a list of N points [(x_1,y_1), (x_2,y_2), ... ] I am trying to find the nearest neighbours to each point based on distance. My dataset is too large to use a brute force approach so a KDtree seems best.

Rather than implement one from scratch I see that sklearn.neighbors.KDTree can find the nearest neighbours. Can this be used to find the nearest neighbours of each particle, i.e return a dim(N) list?

Answer 1

This question is very broad and missing details. It's unclear what you did try, how your data looks like and what a nearest-neighbor is (identity?).

Assuming you are not interested in the identity (with distance 0), you can query the two nearest-neighbors and drop the first column. This is probably the easiest approach here.

Code:

 import numpy as np
 from sklearn.neighbors import KDTree
 np.random.seed(0)
 X = np.random.random((5, 2))  # 5 points in 2 dimensions
 tree = KDTree(X)
 nearest_dist, nearest_ind = tree.query(X, k=2)  # k=2 nearest neighbors where k1 = identity
 print(X)
 print(nearest_dist[:, 1])    # drop id; assumes sorted -> see args!
 print(nearest_ind[:, 1])     # drop id 

Output

 [[ 0.5488135   0.71518937]
  [ 0.60276338  0.54488318]
  [ 0.4236548   0.64589411]
  [ 0.43758721  0.891773  ]
  [ 0.96366276  0.38344152]]
 [ 0.14306129  0.1786471   0.14306129  0.20869372  0.39536284]
 [2 0 0 0 1]

Answer 2

You can use sklearn.neighbors.KDTree's query_radius() method, which returns a list of the indices of the nearest neighbours within some radius (as opposed to returning k nearest neighbours).

from sklearn.neighbors import KDTree

points = [(1, 1), (2, 2), (3, 3), (4, 4), (5, 5)]

tree = KDTree(points, leaf_size=2)
all_nn_indices = tree.query_radius(points, r=1.5)  # NNs within distance of 1.5 of point
all_nns = [[points[idx] for idx in nn_indices] for nn_indices in all_nn_indices]
for nns in all_nns:
    print(nns)

Outputs:

[(1, 1), (2, 2)]
[(1, 1), (2, 2), (3, 3)]
[(2, 2), (3, 3), (4, 4)]
[(3, 3), (4, 4), (5, 5)]
[(4, 4), (5, 5)]

Note that each point includes itself in its list of nearest neighbours within the given radius. If you want to remove these identity points, the line computing all_nns can be changed to:

all_nns = [
    [points[idx] for idx in nn_indices if idx != i]
    for i, nn_indices in enumerate(all_nn_indices)
]

Resulting in:

[(2, 2)]
[(1, 1), (3, 3)]
[(2, 2), (4, 4)]
[(3, 3), (5, 5)]
[(4, 4)]

Answer 3

The sklearn should be the best. I wrote the below some time back ,where I needed custom distance. (I guess sklearn does not support custom distance fn 'KD tree' with custom distance metric . Adding for reference

Adapted from my gist for 2D https://gist.github.com/alexcpn/1f187f2114976e748f4d3ad38dea17e8

# From https://gist.github.com/alexcpn/1f187f2114976e748f4d3ad38dea17e8
# Author alex punnen
from collections import namedtuple
from operator import itemgetter
import numpy as np

def find_nearest_neighbour(node,point,distance_fn,current_axis):
    # Algorith to find nearest neighbour in a KD Tree;the KD tree has done a spatial sort
    # of the given co-ordinates, such that to the left of the root lies co-ordinates nearest to the x-axis
    # and to the right of the root ,lies the co-ordinates farthest from the x axis
    # On the y axis split on the left of the parent/root node lies co-ordinates nearest to the y-axis and to
    # the right of the root, lies the co-ordinates farthest from the y axis
    # to find the nearest neightbour, from the root, you first check left and right node; if distance is closer
    # to the right node,then the entire left node can be discarded from search, because of the spatial split
    # and that node becomes the root node. This process is continued recursively till the nearest is found
    # param:node: The current node
    # param: point: The point to which the nearest neighbour is to be found
    # param: distance_fn: to calculate the nearest neighbour
    # param: current_axis: here assuming only two dimenstion and current axis will be either x or y , 0 or 1

    if node is None:
        return None,None
    current_closest_node = node
    closest_known_distance = distance_fn(node.cell[0],node.cell[1],point[0],point[1])
    print closest_known_distance,node.cell

    x = (node.cell[0],node.cell[1])
    y = point

    new_node = None
    new_closest_distance = None
    if x[current_axis] > y[current_axis]:
        new_node,new_closest_distance= find_nearest_neighbour(node.left_branch,point,distance_fn,
                                                          (current_axis+1) %2)
    else:
        new_node,new_closest_distance = find_nearest_neighbour(node.right_branch,point,distance_fn,
                                                           (current_axis+1) %2) 

    if  new_closest_distance and new_closest_distance < closest_known_distance:
        print 'Reset closest node to ',new_node.cell
        closest_known_distance = new_closest_distance
        current_closest_node = new_node

    return current_closest_node,closest_known_distance


class Node(namedtuple('Node','cell, left_branch, right_branch')):
    # This Class is taken from wikipedia code snippet for  KD tree
    pass

def create_kdtree(cell_list,current_axis,no_of_axis):
    # Creates a KD Tree recursively following the snippet from wikipedia for KD tree
    # but making it generic for any number of axis and changes in data strucure
    if not cell_list:
        return
    # get the cell as a tuple list this is for 2 dimensions
    k= [(cell[0],cell[1])  for cell  in cell_list]
    # say for three dimension
    # k= [(cell[0],cell[1],cell[2])  for cell  in cell_list]
    k.sort(key=itemgetter(current_axis)) # sort on the current axis
    median = len(k) // 2 # get the median of the list
    axis = (current_axis + 1) % no_of_axis # cycle the axis
    return Node(k[median], # recurse 
                create_kdtree(k[:median],axis,no_of_axis),
                create_kdtree(k[median+1:],axis,no_of_axis))

def eucleaden_dist(x1,y1,x2,y2):
    a= np.array([x1,y1])
    b= np.array([x2,y2])
    dist = np.linalg.norm(a-b)
    return dist


np.random.seed(0)
#cell_list = np.random.random((2, 2))
#cell_list = cell_list.tolist()
cell_list = [[2,2],[4,8],[10,2]]
print(cell_list)
tree = create_kdtree(cell_list,0,2)

node,distance = find_nearest_neighbour(tree,(1, 1),eucleaden_dist,0)
print 'Nearest Neighbour=',node.cell,distance

node,distance = find_nearest_neighbour(tree,(8, 1),eucleaden_dist,0)
print 'Nearest Neighbour=',node.cell,distance