Main Content

knnsearch

Find k-nearest neighbors using searcher object

Description

example

Idx = knnsearch(Mdl,Y) searches for the nearest neighbor (i.e., the closest point, row, or observation) in Mdl.X to each point (i.e., row or observation) in the query data Y using an exhaustive search, a Kd-tree, or a Hierarchical Navigable Small Worlds approximate search. knnsearch returns Idx, which is a column vector of the indices in Mdl.X representing the nearest neighbors.

example

Idx = knnsearch(Mdl,Y,Name,Value) returns the indices of the closest points in Mdl.X to Y with additional options specified by one or more Name,Value pair arguments. For example, specify the number of nearest neighbors to search for, or distance metric different from the one stored in Mdl.Distance. You can also specify which action to take if the closest distances are tied.

example

[Idx,D] = knnsearch(___) additionally returns the matrix D using any of the input arguments in the previous syntaxes. D contains the distances between each observation in Y that correspond to the closest observations in Mdl.X. By default, the function arranges the columns of D in ascending order by closeness, with respect to the distance metric.

Examples

collapse all

knnsearch accepts ExhaustiveSearcher or KDTreeSearcher model objects to search the training data for the nearest neighbors to the query data. An ExhaustiveSearcher model invokes the exhaustive searcher algorithm, and a KDTreeSearcher model defines a Kd-tree, which knnsearch uses to search for nearest neighbors.

Load Fisher's iris data set. Randomly reserve five observations from the data for query data.

load fisheriris
rng(1); % For reproducibility
n = size(meas,1);
idx = randsample(n,5);
X = meas(~ismember(1:n,idx),:); % Training data
Y = meas(idx,:);                % Query data

The variable meas contains 4 predictors.

Grow a default four-dimensional Kd-tree.

MdlKDT = KDTreeSearcher(X)
MdlKDT = 
  KDTreeSearcher with properties:

       BucketSize: 50
         Distance: 'euclidean'
    DistParameter: []
                X: [145x4 double]

MdlKDT is a KDTreeSearcher model object. You can alter its writable properties using dot notation.

Prepare an exhaustive nearest neighbor searcher.

MdlES = ExhaustiveSearcher(X)
MdlES = 
  ExhaustiveSearcher with properties:

         Distance: 'euclidean'
    DistParameter: []
                X: [145x4 double]

MdlKDT is an ExhaustiveSearcher model object. It contains the options, such as the distance metric, to use to find nearest neighbors.

Alternatively, you can grow a Kd-tree or prepare an exhaustive nearest neighbor searcher using createns.

Search the training data for the nearest neighbors indices that correspond to each query observation. Conduct both types of searches using the default settings. By default, the number of neighbors to search for per query observation is 1.

IdxKDT = knnsearch(MdlKDT,Y);
IdxES = knnsearch(MdlES,Y);
[IdxKDT IdxES]
ans = 5×2

    17    17
     6     6
     1     1
    89    89
   124   124

In this case, the results of the search are the same.

Grow a Kd-tree nearest neighbor searcher object by using the createns function. Pass the object and query data to the knnsearch function to find k-nearest neighbors.

Load Fisher's iris data set.

load fisheriris

Remove five irises randomly from the predictor data to use as a query set.

rng(1)                      % For reproducibility
n = size(meas,1);           % Sample size
qIdx = randsample(n,5);     % Indices of query data
tIdx = ~ismember(1:n,qIdx); % Indices of training data
Q = meas(qIdx,:);
X = meas(tIdx,:);

Grow a four-dimensional Kd-tree using the training data. Specify the Minkowski distance for finding nearest neighbors.

Mdl = createns(X,'Distance','minkowski')
Mdl = 
  KDTreeSearcher with properties:

       BucketSize: 50
         Distance: 'minkowski'
    DistParameter: 2
                X: [145x4 double]

Because X has four columns and the distance metric is Minkowski, createns creates a KDTreeSearcher model object by default. The Minkowski distance exponent is 2 by default.

Find the indices of the training data (Mdl.X) that are the two nearest neighbors of each point in the query data (Q).

IdxNN = knnsearch(Mdl,Q,'K',2)
IdxNN = 5×2

    17     4
     6     2
     1    12
    89    66
   124   100

Each row of IdxNN corresponds to a query data observation, and the column order corresponds to the order of the nearest neighbors, with respect to ascending distance. For example, based on the Minkowski distance, the second nearest neighbor of Q(3,:) is X(12,:).

Load Fisher's iris data set.

load fisheriris

Remove five irises randomly from the predictor data to use as a query set.

rng(4);                     % For reproducibility
n = size(meas,1);           % Sample size
qIdx = randsample(n,5);     % Indices of query data
X = meas(~ismember(1:n,qIdx),:);
Y = meas(qIdx,:);

Grow a four-dimensional Kd-tree using the training data. Specify the Minkowski distance for finding nearest neighbors.

Mdl = KDTreeSearcher(X);

Mdl is a KDTreeSearcher model object. By default, the distance metric for finding nearest neighbors is the Euclidean metric.

Find the indices of the training data (X) that are the seven nearest neighbors of each point in the query data (Y).

[Idx,D] = knnsearch(Mdl,Y,'K',7,'IncludeTies',true);

Idx and D are five-element cell arrays of vectors, with each vector having at least seven elements.

Display the lengths of the vectors in Idx.

cellfun('length',Idx)
ans = 5×1

     8
     7
     7
     7
     7

Because cell 1 contains a vector with length greater than k = 7, query observation 1 (Y(1,:)) is equally close to at least two observations in X.

Display the indices of the nearest neighbors to Y(1,:) and their distances.

nn5 = Idx{1}
nn5 = 1×8

    91    98    67    69    71    93    88    95

nn5d = D{1}
nn5d = 1×8

    0.1414    0.2646    0.2828    0.3000    0.3464    0.3742    0.3873    0.3873

Training observations 88 and 95 are 0.3873 cm away from query observation 1.

Train two KDTreeSearcher models using different distance metrics, and compare k-nearest neighbors of query data for the two models.

Load Fisher's iris data set. Consider the petal measurements as predictors.

load fisheriris
X = meas(:,3:4); % Predictors
Y = species;     % Response

Train a KDTreeSearcher model object by using the predictors. Specify the Minkowski distance with exponent 5.

KDTreeMdl = KDTreeSearcher(X,'Distance','minkowski','P',5)
KDTreeMdl = 
  KDTreeSearcher with properties:

       BucketSize: 50
         Distance: 'minkowski'
    DistParameter: 5
                X: [150x2 double]

Find the 10 nearest neighbors from X to a query point (newpoint), first using Minkowski then Chebychev distance metrics. The query point must have the same column dimension as the data used to train the model.

newpoint = [5 1.45];
[IdxMk,DMk] = knnsearch(KDTreeMdl,newpoint,'k',10);
[IdxCb,DCb] = knnsearch(KDTreeMdl,newpoint,'k',10,'Distance','chebychev');

IdxMk and IdxCb are 1-by-10 matrices containing the row indices of X corresponding to the nearest neighbors to newpoint using Minkowski and Chebychev distances, respectively. Element (1,1) is the nearest, element (1,2) is the next nearest, and so on.

Plot the training data, query point, and nearest neighbors.

figure;
gscatter(X(:,1),X(:,2),Y);
title('Fisher''s Iris Data -- Nearest Neighbors');
xlabel('Petal length (cm)');
ylabel('Petal width (cm)');
hold on
plot(newpoint(1),newpoint(2),'kx','MarkerSize',10,'LineWidth',2);   % Query point 
plot(X(IdxMk,1),X(IdxMk,2),'o','Color',[.5 .5 .5],'MarkerSize',10); % Minkowski nearest neighbors
plot(X(IdxCb,1),X(IdxCb,2),'p','Color',[.5 .5 .5],'MarkerSize',10); % Chebychev nearest neighbors
legend('setosa','versicolor','virginica','query point',...
   'minkowski','chebychev','Location','Best');

Zoom in on the points of interest.

h = gca; % Get current axis handle.
h.XLim = [4.5 5.5];
h.YLim = [1 2];
axis square;

Several observations are equal, which is why only eight nearest neighbors are identified in the plot.

Create a large data set and compare the speed of an HNSW searcher to the speed of an exhaustive searcher for 1000 query points.

Create the data.

rng default % For reproducibility
A = diag(1:100);
B = randn(100,10);
K = kron(A,B);
ss = size(K)
ss = 1×2

       10000        1000

The data K has 1e4 rows and 1e3 columns.

Create an HNSW searcher object h from the data K.

tic;
h = hnswSearcher(K);
chnsw = toc
chnsw = 10.6544

Create 1000 random query points with 1000 features (columns). Specify to search for five nearest neighbors.

Y = randn(1000, 1000);
tic;
[idx, D] = knnsearch(h,Y,K=5);
thnsw = toc
thnsw = 0.0797

Create an exhaustive searcher object e from the data K.

tic
e = ExhaustiveSearcher(K);
ces = toc
ces = 0.0021

Creating an exhaustive searcher is much faster than creating an HNSW searcher.

Time the search using e and compare the result to the time using the HNSW searcher h.

tic;
[idx0, D0] = knnsearch(e,Y,K=5);
tes = toc
tes = 1.4841

In this case, the HNSW searcher computes about 20 times faster than the exhaustive searcher.

Determine how many results of the HNSW search differ in some way from the results of the exhaustive search.

v = abs(idx - idx0); % Count any difference in the five found entries
vv = sum(v,2); % vv = 0 means no difference
nz = nnz(vv) % Out of 1000 rows how many differ at all?
nz = 118

Here, 118 of 1000 HNSW search results differ from the exhaustive search results.

Try to improve the accuracy of the HNSW searcher by training with nondefault parameters. Specifically, use larger values for MaxNumLinksPerNode and TrainSetSize. These settings affect the speed of training and finding nearest neighbors.

tic;
h2 = hnswSearcher(K,MaxNumLinksPerNode=48,TrainSetSize=2000);
chnsw2 = toc
chnsw2 = 78.4836

With these parameters, the searcher takes about seven times as long to train.

tic;
[idx2, D2] = knnsearch(h2,Y,K=5);
thnsw2 = toc
thnsw2 = 0.1049

The speed of finding nearest neighbors using HNSW decreases, but is still more than ten times faster than the exhaustive search.

v = abs(idx2 - idx0);
vv = sum(v,2);
nz = nnz(vv)
nz = 57

For the slower but more accurate HNSW search, only 57 of 1000 results differ in any way from the exact results. Summarize the timing results in a table.

timet = table([chnsw;ces;chnsw2],[thnsw;tes;thnsw2],'RowNames',["HNSW";"Exhaustive";"HNSW2"],'VariableNames',["Creation" "Search"])
timet=3×2 table
                  Creation      Search 
                  _________    ________

    HNSW             10.654    0.079741
    Exhaustive    0.0021304      1.4841
    HNSW2            78.484     0.10492

Input Arguments

collapse all

Nearest neighbor searcher, specified as an ExhaustiveSearcher, KDTreeSearcher, or hnswSearcher model object.

If Mdl is an ExhaustiveSearcher model, then knnsearch searches for nearest neighbors using an exhaustive search. If Mdl is a KDTreeSearcher model, knnsearch uses the grown Kd-tree to search for nearest neighbors. If Mdl is an hnswSearcher model, knnsearch uses the Hierarchical Navigable Small Worlds approximate neighbor search algorithm. For descriptions, see k-Nearest Neighbor Search and Radius Search.

Query data, specified as a numeric matrix.

Y is an m-by-K matrix. Rows of Y correspond to observations (i.e., examples), and columns correspond to predictors (i.e., variables or features). Y must have the same number of columns as the training data stored in Mdl.X.

Data Types: single | double

Name-Value Arguments

Specify optional pairs of arguments as Name1=Value1,...,NameN=ValueN, where Name is the argument name and Value is the corresponding value. Name-value arguments must appear after other arguments, but the order of the pairs does not matter.

Before R2021a, use commas to separate each name and value, and enclose Name in quotes.

Example: 'K',2,'Distance','minkowski' specifies to find the two nearest neighbors of Mdl.X to each point in Y and to use the Minkowski distance metric.

For Exhaustive Nearest Neighbor Searchers and Kd-Tree Nearest Neighbor Searchers

collapse all

Distance metric used to find neighbors of the training data to the query observations, specified as one of the values in this table or function handle.

For all types of nearest neighbor searchers, knnsearch supports these distance metrics.

ValueDescription
'chebychev'Chebychev distance (maximum coordinate difference)
'cityblock'City block distance
'euclidean'Euclidean distance
'minkowski'Minkowski distance. The default exponent is 2. To specify a different exponent, use the 'P' name-value argument.

If Mdl is an ExhaustiveSearcher model object, then knnsearch also supports these distance metrics.

ValueDescription
'correlation'One minus the sample linear correlation between observations (treated as sequences of values)
'cosine'One minus the cosine of the included angle between observations (treated as row vectors)
'fasteuclidean'Euclidean distance computed by using an alternative algorithm that saves time when the number of predictors is at least 10. In some cases, this faster algorithm can reduce accuracy. This distance metric is available only when NSMethod is 'exhaustive'. Algorithms starting with 'fast' do not support sparse data. For details, see Algorithms.
'fastseuclidean'Standardized Euclidean distance computed by using an alternative algorithm that saves time when the number of predictors is at least 10. In some cases, this faster algorithm can reduce accuracy. This distance metric is available only when NSMethod is 'exhaustive'. Algorithms starting with 'fast' do not support sparse data. For details, see Algorithms.
'hamming'Hamming distance, which is the percentage of coordinates that differ
'jaccard'One minus the Jaccard coefficient, which is the percentage of nonzero coordinates that differ
'mahalanobis'Mahalanobis distance, computed using a positive definite covariance matrix. To change the value of the covariance matrix, use the 'Cov' name-value argument.
'seuclidean'Standardized Euclidean distance. Each coordinate difference between the rows in X and the query matrix Y is scaled by dividing by the corresponding element of the standard deviation computed from X. To specify a different scaling, use the 'Scale' name-value argument.
'spearman'One minus the sample Spearman's rank correlation between observations (treated as sequences of values)

If Mdl is an hnswSearcher model object, knnsearch supports all the distances in the ExhaustiveSearcher table except for those beginning with fast: "fasteuclidean" and "fastseuclidean".

If Mdl is an ExhaustiveSearcher model object, then you can also specify a function handle for a custom distance metric by using @ (for example, @distfun). The custom distance function must:

  • Have the form function D2 = distfun(ZI,ZJ).

  • Take as arguments:

    • A 1-by-K vector ZI containing a single row from Mdl.X or Y, where K is the number of columns of Mdl.X.

    • An m-by-K matrix ZJ containing multiple rows of Mdl.X or Y, where m is a positive integer.

  • Return an m-by-1 vector of distances D2, where D2(j) is the distance between the observations ZI and ZJ(j,:).

For more details, see Distance Metrics.

Example: 'Distance','minkowski'

Data Types: char | string | function_handle

Flag to include nearest neighbors that have the same distance from query observations, specified as the comma-separated pair consisting of 'IncludeTies' and false (0) or true (1).

If IncludeTies is true, then:

  • knnsearch includes all nearest neighbors whose distances are equal to the kth smallest distance in the output arguments, where k is the number of searched nearest neighbors specified by the 'K' name-value pair argument.

  • Idx and D are m-by-1 cell arrays such that each cell contains a vector of at least k indices and distances, respectively. Each vector in D contains arranged distances in ascending order. Each row in Idx contains the indices of the nearest neighbors corresponding to the distances in D.

If IncludeTies is false, then knnsearch chooses the observation with the smallest index among the observations that have the same distance from a query point.

Example: 'IncludeTies',true

Number of nearest neighbors to search for in the training data per query observation, specified as the comma-separated pair consisting of 'K' and a positive integer.

Example: 'K',2

Data Types: single | double

Exponent for the Minkowski distance metric, specified as a positive scalar. This argument is valid only when Distance is "minkowski".

The value of P sets the value of the DistParameter property in the model object.

Example: P=3

Data Types: single | double

For Kd-Tree Nearest Neighbor Searchers

collapse all

Flag to sort returned indices according to distance, specified as the comma-separated pair consisting of 'SortIndices' and either true (1) or false (0).

For faster performance, you can set SortIndices to false when the following are true:

  • Y contains many observations that have many nearest neighbors in X.

  • Mdl is a KDTreeSearcher model object.

  • IncludeTies is false.

In this case, knnsearch returns the indices of the nearest neighbors in no particular order. When SortIndices is true, the function arranges the nearest neighbor indices in ascending order by distance.

SortIndices is true by default. When Mdl is an ExhaustiveSearcher model object or IncludeTies is true, the function always sorts the indices.

Example: 'SortIndices',false

Data Types: logical

For Exhaustive Nearest Neighbor Searchers

collapse all

Covariance matrix for the Mahalanobis distance metric, specified as the comma-separated pair consisting of 'Cov' and a positive definite matrix. Cov is a K-by-K matrix, where K is the number of columns of Mdl.X. If you specify Cov and do not specify 'Distance','mahalanobis', then knnsearch returns an error message.

Example: 'Cov',eye(3)

Data Types: single | double

Scale parameter value for the standardized Euclidean distance metric, specified as the comma-separated pair consisting of 'Scale' and a nonnegative numeric vector. Scale has length K, where K is the number of columns of Mdl.X.

The software scales each difference between the training and query data using the corresponding element of Scale. If you specify Scale and do not specify 'Distance','seuclidean', then knnsearch returns an error message.

Example: 'Scale',quantile(Mdl.X,0.75) - quantile(Mdl.X,0.25)

Data Types: single | double

For HNSW Searchers

collapse all

Size of the candidate list of nearest neighbors for a single query point during the search process, specified as a positive integer from K through N. Larger values lead to a more complete search, which increases the likelihood of finding the true nearest neighbors, but slow the search. SearchSetSize must be at least K, the number of features in the data (number of columns in Mdl.X), and must be no more than N, the number of rows in the training data (number of rows in Mdl.X).

Example: SearchSetSize=15

Data Types: double

Note

If you specify 'Distance', 'Cov', 'P', or 'Scale', then Mdl.Distance and Mdl.DistParameter do not change value.

Output Arguments

collapse all

Training data indices of nearest neighbors, returned as a numeric matrix or cell array of numeric vectors.

  • If you do not specify IncludeTies (false by default), then Idx is an m-by-k numeric matrix, where m is the number of rows in Y and k is the number of searched nearest neighbors specified by the 'K' name-value pair argument. Idx(j,i) indicates that Mdl.X(Idx(j,i),:) is one of the k closest observations in Mdl.X to the query observation Y(j,:).

  • If you specify 'IncludeTies',true, then Idx is an m-by-1 cell array such that cell j (Idx{j}) contains a vector of at least k indices of the closest observations in Mdl.X to the query observation Y(j,:).

If SortIndices is true, then knnsearch arranges the indices in ascending order by distance.

Distances of the nearest neighbors to the query data, returned as a numeric matrix or cell array of numeric vectors.

  • If you do not specify IncludeTies (false by default), then D is an m-by-k numeric matrix, where m is the number of rows in Y and k is the number of searched nearest neighbors specified by the 'K' name-value pair argument. D(j,i) is the distance between Mdl.X(Idx(j,i),:) and the query observation Y(j,:) with respect to the distance metric.

  • If you specify 'IncludeTies',true, then D is an m-by-1 cell array such that cell j (D{j}) contains a vector of at least k distances of the closest observations in Mdl.X to the query observation Y(j,:).

If SortIndices is true, then knnsearch arranges the distances in ascending order.

Tips

knnsearch finds the k (positive integer) points in Mdl.X that are k-nearest for each Y point. In contrast, rangesearch finds all the points in Mdl.X that are within distance r (positive scalar) of each Y point.

Algorithms

collapse all

Fast Euclidean Distance Algorithm

The values of the Distance argument that begin fast (such as 'fasteuclidean' and 'fastseuclidean') calculate Euclidean distances using an algorithm that uses extra memory to save computational time. This algorithm is named "Euclidean Distance Matrix Trick" in Albanie [1] and elsewhere. Internal testing shows that this algorithm saves time when the number of predictors is at least 10. Algorithms starting with 'fast' do not support sparse data.

To find the matrix D of distances between all the points xi and xj, where each xi has n variables, the algorithm computes distance using the final line in the following equations:

Di,j2=xixj2=(xixj)T(xixj)=xi22xiTxj+xj2.

The matrix xiTxj in the last line of the equations is called the Gram matrix. Computing the set of squared distances is faster, but slightly less numerically stable, when you compute and use the Gram matrix instead of computing the squared distances by squaring and summing. For a discussion, see Albanie [1].

References

[1] Albanie, Samuel. Euclidean Distance Matrix Trick. June, 2019. Available at https://www.robots.ox.ac.uk/%7Ealbanie/notes/Euclidean_distance_trick.pdf.

Alternative Functionality

  • knnsearch is an object function that requires an ExhaustiveSearcher, KDTreeSearcher, or hnswSearcher model object and query data. Under equivalent conditions for an exhaustive or Kd-tree search, the knnsearch object function returns the same results as the knnsearch function with the specified name-value argument NSMethod="exhaustive" or NSMethod="kdtree", respectively. Note that the knnsearch function does not provide a similar name-value argument for specifying an hnswSearcher model.

  • For k-nearest neighbors classification, see fitcknn and ClassificationKNN.

References

[1] Friedman, J. H., Bentley, J., and Finkel, R. A. (1977). “An Algorithm for Finding Best Matches in Logarithmic Expected Time.” ACM Transactions on Mathematical Software Vol. 3, Issue 3, Sept. 1977, pp. 209–226.

[2] Malkov, Yu. A., and D. A. Yashunin. Efficient and Robust Approximate Nearest Neighbor Search Using Hierarchical Navigable Small World Graphs. Available at https://arxiv.org/abs/1603.09320.

Extended Capabilities

Version History

Introduced in R2010a

expand all