regression same values predicted

[karasjoh000]

When I train my regression problem, I only get 1 and -1 for the weights of support vectors and it produces an exact same prediction for every test vectors.

[cjlin1]

Could you give more details? What's the screen output of the training process?

…

On 2019-03-18 05:40, John Karasev wrote: When I train my regression problem, I only get 1 and -1 for the weights of support vectors and it produces an exact same prediction for the test vectors. -- You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub [1], or mute the thread [2]. [ { ***@***.***": "http://schema.org", ***@***.***": "EmailMessage", "potentialAction": { ***@***.***": "ViewAction", "target": "#139", "url": "#139", "name": "View Issue" }, "description": "View this Issue on GitHub", "publisher": { ***@***.***": "Organization", "name": "GitHub", "url": "https://github.com" } } ] Links: ------ [1] #139 [2] https://github.com/notifications/unsubscribe-auth/AFGwnij8M33tZJm7v3xaAjZJZhbS7Ppxks5vX4k4gaJpZM4b5rop

[Nuist-Lee]

What I got is the same !
I always get the same prediction value, even if I've already changed the inputdata.
Specifically, I used 9 precipitation factors and real precipitation data from 1951 to 2010 to train the SVM regression model. Then I input the factors from 2011 to 2016 to the model in order to get the predicted precipitation data. But all I got was an array with the same 6 numberslike[248 248 248 248 248 248]. When I change the inputdata, the result may change but still an array with the same six numbers. Sometimes the result array can be 'almost' the same, like [222.121971040339 217.475177855453 221.321911068434 215.667836808570 222.829461145174].
It's also been found that only if I input the data that is very very closed to the trainnig data(1951-2010 factors), the result can be quite well.
Ironically, when I use some classic dataset (from my reference books), the result will be pretty pretty good.
Add on, RBF kernel was used, too.

[cjlin1]

Could you provide more details? Sometimes you need to scale the target values before training SVR.

…

On 2022-04-28 00:13, Nuist-Lee wrote: What I got is the same ! I always get the same prediction value, even if I've already changed the inputdata. Specifically, I used 9 precipitation factors and real precipitation data from 1951 to 2010 to train the SVM regression model. Then I input the factors from 2011 to 2016 to the model in order to get the predicted precipitation data. But ALL I GOT WAS AN ARRAY WITH THE SAME 6 NUMBERSlike[248 248 248 248 248 248]. When I change the inputdata, the result may change but still an array with the same six numbers. Sometimes the result array can be 'almost' the same, like [222.121971040339 217.475177855453 221.321911068434 215.667836808570 222.829461145174]. It's also been found that only if I input the data that is VERY VERY closed to the trainnig data(1951-2010 factors), the result can be quite well. Ironically, when I use some classic dataset (from my reference books), the result will be pretty pretty good. Add on, RBF kernel was used, too. -- Reply to this email directly, view it on GitHub [1], or unsubscribe [2]. You are receiving this because you commented.Message ID: ***@***.***> [ { ***@***.***": "http://schema.org", ***@***.***": "EmailMessage", "potentialAction": { ***@***.***": "ViewAction", "target": "#139 (comment)", "url": "#139 (comment)", "name": "View Issue" }, "description": "View this Issue on GitHub", "publisher": { ***@***.***": "Organization", "name": "GitHub", "url": "https://github.com" } } ] Links: ------ [1] #139 (comment) [2] https://github.com/notifications/unsubscribe-auth/ABI3BHWTXT32F7WA54ORUZTVHFRRZANCNFSM4G7GXIUQ

[Nuist-Lee]

This is the code for trainning.

clear all
load ("predictionfactors.mat");

[tsx,TSXps]=mapminmax(InputdataS1');
[ts,TSps]=mapminmax(JHy');
TSX=tsx';
TS=ts';


%% 选择回归预测分析最佳的SVM参数c&g

[bestmse,bestc,bestg] = gaSVMcgForRegress(TS,TSX);


% 打印选择结果
disp('打印选择结果');
str = sprintf( 'Best Cross Validation MSE = %g Best c = %g Best g = %g',bestmse,bestc,bestg);
disp(str);

%% 利用回归预测分析最佳的参数进行SVM网络训练
cmd = ['-c ', num2str(bestc), ' -g ', num2str(bestg) , ' -s 3 -p 0.01 -t 2 '];
model = svmtrain(TS,TSX,cmd);

When I load ("predictionfactors.mat"), the workspace is as follows:

InputdataS1 is precipitation factors in the former 60 years, which is the trainning data.InputdataS2 is that in the latter 6 years, and is used for prediction.
JHy is precipitation data in the former 60 years, and JHt is the former one.
gaSVMcgForRegressis a function written by Faruto, which can find the best c and g in SVM by using Genetic Algorithm.
Then is the code for SVM regression prediction:

    ttx=mapminmax('apply',InputdataS2',TSXps);
    tty=mapminmax('apply',JHt',TSps);
    TTX=ttx';
    TTY=tty';
    
    [predict_,mse] = svmpredict(TTY,TTX,model);
    predict_ = mapminmax('reverse',predict_',TSps);
    predict_ = predict_';

And then the value 'predict_' always be an array with same numbers, which means however I change my inputdata to the model, the prediction result is the same one.

[Nuist-Lee]

In addition , @kevin031060 said he met the same issue when using Java interface ,and Python/Matlab is fine.While mine is using Matlab toolbox.

[cjlin1]

how does your TSX and TS look like? Could you show a few elements of them?

…

On 2022-04-28 13:26, Nuist-Lee wrote: This is the code for trainning. clear all load ("predictionfactors.mat"); [tsx,TSXps]=mapminmax(InputdataS1'); [ts,TSps]=mapminmax(JHy'); TSX=tsx'; TS=ts'; %% 选择回归预测分析最佳的SVM参数c&g [bestmse,bestc,bestg] = gaSVMcgForRegress(TS,TSX); % 打印选择结果 disp('打印选择结果'); str = sprintf( 'Best Cross Validation MSE = %g Best c = %g Best g = %g',bestmse,bestc,bestg); disp(str); %% 利用回归预测分析最佳的参数进行SVM网络训练 cmd = ['-c ', num2str(bestc), ' -g ', num2str(bestg) , ' -s 3 -p 0.01 -t 2 ']; model = svmtrain(TS,TSX,cmd); When I load ("predictionfactors.mat"), the workspace is as follows: [1] InputdataS1 is precipitation factors in the former 60 years, which is the trainning data.InputdataS2 is that in the latter 6 years, and is used for prediction. JHy is precipitation data in the former 60 years, and JHt is the former one. GASVMCGFORREGRESSis a function written by Faruto, which can find the best c and g in SVM by using Genetic Algorithm. Then is the code for SVM regression prediction: ttx=mapminmax('apply',InputdataS2',TSXps); tty=mapminmax('apply',JHt',TSps); TTX=ttx'; TTY=tty'; [predict_,mse] = svmpredict(TTY,TTX,model); predict_ = mapminmax('reverse',predict_',TSps); predict_ = predict_'; And then the value 'predict_' always be an array with same numbers, which means however I change my inputdata to the model, the prediction result is the same one. -- Reply to this email directly, view it on GitHub [2], or unsubscribe [3]. You are receiving this because you commented.Message ID: ***@***.***> [ { ***@***.***": "http://schema.org", ***@***.***": "EmailMessage", "potentialAction": { ***@***.***": "ViewAction", "target": "#139 (comment)", "url": "#139 (comment)", "name": "View Issue" }, "description": "View this Issue on GitHub", "publisher": { ***@***.***": "Organization", "name": "GitHub", "url": "https://github.com" } } ] Links: ------ [1] https://user-images.githubusercontent.com/78215393/165681156-c3ffb18d-82bd-4c3b-9981-0da5b3189034.png [2] #139 (comment) [3] https://github.com/notifications/unsubscribe-auth/ABI3BHROA6QBSFC5FZSJQU3VHIOQBANCNFSM4G7GXIUQ

[Nuist-Lee]

TSX is prediction factors which are already transposed and normalizd and transposed again to be a 601 matrix.
TS is real precipitation data, operated as above. It's a 609 matrix.

[Nuist-Lee]

I mean 60x1 and 60x9

[Nuist-Lee]

问题解决了！谢谢林老师的耐心指导！
Issue solved!!! Thanks for your patient help!!!
The key is Y, which is the real precipitation data in my code named TS.
In fact , I have 3 precipitation data for 3 different regions. And when I choose JH region, the issue occurs.
I was always focused on the X, which is prediction factor in my code. While the Y is another part of trainning an SVR model, which was ignored for a long time.
So the result is : One region cannot match well with those precipitation factors, while another region can.

[cjlin1]

OK. Glad to know your problems are solved

…

On 2022-04-29 01:18, Nuist-Lee wrote: 问题解决了！谢谢林老师的耐心指导！ Issue solved!!! Thanks for your patient help!!! The key is Y, which is the real precipitation data in my code named TS. In fact , I have 3 precipitation data for 3 different regions. And when I choose JH region, the issue occurs. I was always focused on the X, which is prediction factor in my code. While the Y is another part of trainning an SVR model, which was ignored for a long time. So the result is : One region cannot match well with those precipitation factors, while another region can. -- Reply to this email directly, view it on GitHub [1], or unsubscribe [2]. You are receiving this because you commented.Message ID: ***@***.***> [ { ***@***.***": "http://schema.org", ***@***.***": "EmailMessage", "potentialAction": { ***@***.***": "ViewAction", "target": "#139 (comment)", "url": "#139 (comment)", "name": "View Issue" }, "description": "View this Issue on GitHub", "publisher": { ***@***.***": "Organization", "name": "GitHub", "url": "https://github.com" } } ] Links: ------ [1] #139 (comment) [2] https://github.com/notifications/unsubscribe-auth/ABI3BHQBNKWNO2ARFYCP2RLVHLB6TANCNFSM4G7GXIUQ