fastknn - WeAreCAS

Q: What is the primary function of the fastknn action?

The fastknn action performs a k-nearest neighbor search using the specified input tables.

Q: Which input tables are required for the fastknn action?

The action requires the 'query' parameter, which specifies the query input data table containing query observations, and the 'table' parameter, which specifies the settings for the input table to search against.

Q: What are the available distance metrics for the k-nearest neighbor calculation?

The 'distanceMetric' parameter supports 'COSINE' (cosine distance), 'IP' (inner product), and 'L2' (Euclidean distance). The default value is 'L2'.

Q: How can missing values in the query data be handled?

By setting the 'impute' parameter to TRUE, observations with missing values in the query data table can be imputed using the k-nearest-neighbors method. The default is FALSE.

Q: What search methods are available in the fastknn action?

The 'method' parameter allows you to choose between 'APPROXIMATE' (approximate search) and 'EXACT' (exact search). The default method is 'EXACT'.

Q: What does the efConstruction parameter control?

The 'efConstruction' parameter specifies the number of neighbors to consider during graph construction. The default value is 200, and the minimum is 1.

Q: How can I specify the number of neighbors to return?

You can specify the number of neighbors to be returned using the 'k' parameter. The default value is 2.

Q: What output tables can be created by the fastknn action?

The action can create an 'outDist' table (for computed distances), an 'outImpute' table (for query data after imputation), and an 'output' table (casOut) for the computed neighbors.

Q: What is the purpose of the useTopKOutDist parameter?

When set to TRUE, the 'useTopKOutDist' parameter specifies that only the top k nearest distances should be output to the outDist parameter table.

Description

The fastknn action performs a k-nearest neighbor search. It calculates the distances between observations in a query table and an input table to identify the k-nearest neighbors. It supports both exact and approximate search methods using various distance metrics (Euclidean, Cosine, Inner Product) and allows for missing value imputation.

Settings

Parameter	Description
table	Specifies the settings for the input table containing the data to be searched.
query	Specifies the query input data table containing the observations for which to find neighbors.
k	Specifies the number of nearest neighbors to be returned. The default is 2.
method	Specifies the k-nearest neighbor search method to use. Values can be 'EXACT' (default) or 'APPROXIMATE'.
distanceMetric	Specifies the metric to measure the distance between points. Options are 'L2' (Euclidean, default), 'COSINE', or 'IP' (Inner Product).
inputs	Specifies the variables to use in the analysis for distance calculation.
id	Specifies one or more variables to use as record identifiers.
output	Specifies the output data table in which to save the computed neighbors.
outDist	Specifies the output data table in which to save the computed distances.
outImpute	Specifies the output data table in which to save the query data after imputing missing values.
impute	When set to True, enables imputation of missing values in the query table using the k-nearest neighbors method.
efConstruction	Specifies the number of neighbors to consider during graph construction (for approximate search).
efSearch	Specifies the number of candidate nodes to explore during the graph search phase (for approximate search).

Data Preparation View data prep sheet

Create Input and Query Data

Create two datasets: one serving as the reference data and one as the query data.

Copied!

1	DATA mycas.inputData; INPUT id x y; CARDS; 1 1 1
2	2 1 2
3	3 2 1
4	4 2 2
5	5 5 5
6	6 5 6
7	7 6 5
8	8 6 6
9	; RUN;
10
11	DATA mycas.queryData; INPUT id x y; CARDS; 1 1.5 1.5
12	2 5.5 5.5
13	; RUN;

Examples

Perform an exact search to find the 2 nearest neighbors for each query point using Euclidean distance.

SAS® / CAS Code Code awaiting community validation

Copied!

1
2	PROC CAS;
3	fastKnn.fastknn / TABLE={name='inputData'} query={name='queryData'} k=2 id={'id'} inputs={'x', 'y'} OUTPUT={casOut={name='knn_results', replace=true}};
4
5	RUN;
6

Result :
The 'knn_results' table will contain the 2 nearest neighbors from 'inputData' for each observation in 'queryData'.

Perform an approximate search for 3 neighbors using the Cosine distance metric, and output the calculated distances to a separate table.

SAS® / CAS Code Code awaiting community validation

Copied!

1
2	PROC CAS;
3	fastKnn.fastknn / TABLE={name='inputData'} query={name='queryData'} k=3 method='APPROXIMATE' distanceMetric='COSINE' id={'id'} inputs={'x', 'y'} outDist={name='dist_results', replace=true} OUTPUT={casOut={name='knn_approx_results', replace=true}};
4
5	RUN;
6

Result :
Two output tables are created: 'knn_approx_results' containing the neighbors and 'dist_results' containing the distances between the query points and their neighbors.

FAQ

What is the primary function of the fastknn action?

Which input tables are required for the fastknn action?

What are the available distance metrics for the k-nearest neighbor calculation?

How can missing values in the query data be handled?

What search methods are available in the fastknn action?

What does the efConstruction parameter control?

How can I specify the number of neighbors to return?

What output tables can be created by the fastknn action?

What is the purpose of the useTopKOutDist parameter?

Table of Contents

Description

Create Input and Query Data

Examples

Simple Exact k-Nearest Neighbor Search

Approximate Search with Distance Output

FAQ