# How to use nearest neighbor to find similar population based on list of features

Nearest neighbor can classify new data point based on the k nearest neighbor's class. Assuming there is dataset A contains 10000 data points. There is also another dataset B contains 1 MM data points. The goal is to find the most similar records from dataset B that resembles dataset A on a number of pre-decided attributes(features). Assume there is a list of specific features that I'm interested in, calculate a distance between the one record and all the other records and pick the records with the smallest distance can serve this purpose.

SAS has a couple of procedure can do nearest neighbor such as PROC DISCRIM that takes a training data and classify on the test data such as below. In this case, how to define training data as the purpose is just to find the most similar records in dataset B that looks like each individual records in data A? Can I construct a training data by randomly take 50% of dataset B and combine with dataset A as training data, and the rest 50% of dataset B as test data?

proc discrim data=train method=npar k=5 testdata=toscore testout=toscore_out ; class y; var x1-x10; /* a list of features to compare */ run;

• Can you clarify on a simple example what exactly you want to achieve? (data A; input a b c; cards; ...; run; data B; input a b c; cards; ...; run;) Is it that for each individual record of data set A you want to find its most similar record in data set B? – hannafrc May 10 '17 at 12:45
• Yes. for each individual record of dataset A, need to find its most similar record in data set B based on a list of pre-defined attributes in dataset A. – Eric_IL May 13 '17 at 3:09

Regarding proc discrim - considering PROC DISCRIM documentation I don't see a possibility to get for each observation from test data the closest neighbour (observation) from train data. I've found a similar problem described here: https://stackoverflow.com/questions/19626326/k-nearest-neighbor-in-sas-how-to-get-the-neighbor-list-for-each-row. Also without answer.

I'd suggest to use another procedure. Here is an example workaround with the use of proc modeclus. However, it needs further work, as for each B's (testing set's) record it loops through the whole table A (training set). It also makes no use of y label from table A.

1.Example data

        data A; input a b c y; cards;
1 1 1 1
1 1 2 1
2 2 2 2
2 2 3 2
4 4 4 3
4 4 5 3
;
run;
data B; input a b c; cards;
1 1 4
2 2 3
3 3 3
3 3 4
;
run;

1. Find nearest neighbors of A in B. Put results to table C.

    %macro findNN(A,B,C);
/* get table sizes */
%let size1=;
proc sql;
select count(*) into :size1
from &B.;
select count(*) into :size2
from &A.;
quit;
%let size2=%eval(&size2.+1);
/* for each observation from table B*/
%do i=1 %to &size1.;
data AB;
set &A. &B.(firstobs=&i. obs=&i.);
keep a b c;
run;
/* find its nearest neighbour from table A */
ods select Neighbor;
proc modeclus data=AB method=1 k=2 Neighbor;
var a b c;
ods output Neighbor=tableout;
run;
/* add the neighbor found to the table C */
%if &i.=1 %then %do;
data &C.;
set tableout(where=(compress(id)="&size2.") keep=id nbor distance in=out);
if out=1 then id=&i.;
run;
%end;
%else %do;
data &C.;
set &C. tableout(where=(compress(id)="&size2.") keep=id nbor distance in=out);
if out=1 then id=&i.;
run;
%end;
%end;
%mend findNN;
%findNN(A,B,ANborB);