Optimize Fitted KNN Classifier

Question

Hyeongcheol Lee 2019-11-20

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/492050-optimize-fitted-knn-classifier

评论： Hyeongcheol Lee 2019-11-28

I have questions about the "Optimize Fitted KNN Classifier" example (https://www.mathworks.com/help/stats/fitcknn.html).

Is there any rule to change values of "NumNeighbors" and "Distance" while iteration?

It just seems random.

And, I'm not sure the final model with below "NumNeighbors" and "Distance" is the best optimized model or not.

How can you assure the final model is the best optimized?

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Bhargavi Maganuru 2019-11-28

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/492050-optimize-fitted-knn-classifier#answer_403823

Values of “NumNeighbors” and “Distance” are changing because of setting property ‘OptimizeHyperparameters' to ‘auto’ which will try to optimize the “distance” and “NumNeighbours” parameters.

Refer to the following link for more information about ‘OptimizeHyperparameters'

https://www.mathworks.com/help/stats/fitcknn.html#d117e332170

To pick best estimate, bayesian optimization acquisition function 'expected-improvement-plus' is used. It calculates best estimated feasible point using ‘bestpoint’ function.

For more information about ‘bestpoint’ refer the link

https://www.mathworks.com/help/stats/bayesianoptimization.bestpoint.html

Hope this helps!

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Hyeongcheol Lee 2019-11-28

Hello, Bhargavi.

Thanks for your reply.

But, I'm still not sure about below two things.

Is there any rule to change values of "NumNeighbors" and "Distance" while iteration? In other words, is there any rule to choose "Next point"?

2. In the example "Optimize Fitted KNN Classifier" in the below site, the iteration is only 30. So, if we do more iteration I think there is better optimization. So, my question is... how should I do to get real optimization point?

https://www.mathworks.com/help/stats/fitcknn.html#d117e332170

请先登录，再进行评论。