Clear Filters
Clear Filters

In the following code, Why is the classification accuracy (acc1, acc2) calculated differently?

3 views (last 30 days)
The following function (knn_test) takes the following parameters: X indicates dataset sampels, Y indicates dataset labels, and filterindex corresponds column filter. If I want to filter the data, filterIndex is set to 1, whichever column is to be filtered.
I want to validate two holdout method. One with crossval function, and the other with cvpartition function. But when I call this method, acc1 and acc2 variables show different values.
I added break point to the code and debugged the code. I examined CVKNNModels' partition property and it was the same with the c particion in the Model 2.
What could have gone wrong with the following code?
Why did these two accuracy variables take different values?
If I want to use this function for holdout classification, which model should I choose?
Thanks.
function [acc1,acc2]=knn_test(X,Y,filterIndex)
columnfilterIndex = find(filterIndex==1);
%Model 1
tra1Data = X(:,[columnfilterIndex]);
tra1Label=Y;
KNNModel1=fitcknn(tra1Data, tra1Label, 'Distance', 'Euclidean', 'NumNeighbors', 3, 'DistanceWeight', 'Equal', 'Standardize', true);
rng('default');
CVKNNModel = crossval(KNNModel1,'holdout',0.3);
loss=kfoldLoss(CVKNNModel);
acc1=1-loss;
%Model 2
rng('default');
c = cvpartition(Y,'HoldOut',0.3);
tra2Data=X(c.training,[columnfilterIndex]);
tra2Label=Y(c.training,:);
test2Data=X(c.test,[columnfilterIndex]);
test2Label=Y(c.test,:);
KNNModel2 = fitcknn(tra2Data,tra2Label,'Distance', 'Euclidean','NumNeighbors',3, 'DistanceWeight', 'Equal','Standardize', true);
pre_test = predict(KNNModel2,test2Data);
correctPredictions = (pre_test == test2Label);
acc2 = sum(correctPredictions)/length(correctPredictions);
%perf=classperf(uint8(test2Label),uint8(pre_test));
%acc2=perf.CorrectRate;
  2 Comments
michio
michio on 12 Nov 2019
Could you provide a script that can reproduce the issue? I run the following and acc1 and acc2 are the same.
load ionosphere
[acc1,acc2]=knn_test(X,Y,ones(size(X,2),1))
where (note the line: correctPredictions = (string(pre_test) == string(test2Label)); to avoid error)
function [acc1,acc2]=knn_test(X,Y,filterIndex)
columnfilterIndex = find(filterIndex==1);
%Model 1
tra1Data = X(:,[columnfilterIndex]);
tra1Label=Y;
KNNModel1=fitcknn(tra1Data, tra1Label, 'Distance', 'Euclidean', 'NumNeighbors', 3, 'DistanceWeight', 'Equal', 'Standardize', true);
rng('default');
CVKNNModel = crossval(KNNModel1,'holdout',0.3);
loss=kfoldLoss(CVKNNModel);
acc1=1-loss;
%Model 2
rng('default');
c = cvpartition(Y,'HoldOut',0.3);
tra2Data=X(c.training,[columnfilterIndex]);
tra2Label=Y(c.training,:);
test2Data=X(c.test,[columnfilterIndex]);
test2Label=Y(c.test,:);
KNNModel2 = fitcknn(tra2Data,tra2Label,'Distance', 'Euclidean','NumNeighbors',3, 'DistanceWeight', 'Equal','Standardize', true);
pre_test = predict(KNNModel2,test2Data);
% correctPredictions = (pre_test == test2Label);
correctPredictions = (string(pre_test) == string(test2Label));
acc2 = sum(correctPredictions)/length(correctPredictions);
%perf=classperf(uint8(test2Label),uint8(pre_test));
%acc2=perf.CorrectRate;
Y. K.
Y. K. on 13 Nov 2019
Edited: Y. K. on 14 Nov 2019
I couldn't add test data. Because it size exceeds 5mb. I send the data to you with e-mail.
Thanks for your interest.

Sign in to comment.

Accepted Answer

michio
michio on 18 Nov 2019
Edited: michio on 18 Nov 2019
The two ways of hold-out cross-validation that you described have some subtle differences. For Model 1, when calling the crossval method on the KNNModel1, the prior is based on the whole dataset. For Model 2, the prior is based on the training partition tra2Data. If you specify the same prior, you should get the same results.
KNNModel2 = fitcknn(tra2Data,tra2Label,...
'Distance', 'Euclidean','NumNeighbors',3, ...
'DistanceWeight', 'Equal','Standardize', true, 'Prior', KNNModel1.Prior);

More Answers (1)

Y. K.
Y. K. on 12 Nov 2019
In addition, this code gave the same results for small datasets. But when the number of features increases, it produces different results, especially in my dataset.

Products


Release

R2019b

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!