matlab - Decision Level Fusion of SVR outputs -

i have 2 sets of features predicting same outputs. instead of training @ once, train them separately , fuse decisions. in svm classification, can take probability values classes can used train svm. in svr, how can this?

any ideas?

thanks :)

there couple of choices here . 2 popular ones be:

one)

build 2 models , average results.

it tends work in practice.

two)

you in similar fashion when have probabilities. problem is, need control on fitting .what mean is "dangerous" produce score 1 set of features , apply labels same before (even if new features different). because new applied score trained on these labels , therefore on fits in (hyper-performs).

normally use cross-validation

in case have

1. train_set_1 x1 features , label y
2. train_set_2 x2 features , same label y

some psedo code:

randomly split 50-50 both train_set_1 , train_set_2 @ same points along y (output array) 

so have:

a.train_set_1 (50% of training_set_1) b.train_set_1 (the rest of 50% of training_set_1) a.train_set_2 (50% of training_set_2) b.train_set_2 (the rest of 50% of training_set_2) a.y (50% of output array corresponds same sets a.train_set_1 , a.train_set_2) b.y (50% of output array corresponds same sets b.train_set_1 , b.train_set_2) 

here key part

build svr a.train_set_1 (that contains x1 features) , output a.y ,  apply model's prediction feature b.train_set_2 .  mean, score b.train_set_2 base on first model. take score , paste next a.train_set_2 .so set have x2 features + 1 more feature, score produced first model.  build final model on b.train_set_2 , b.y  

the new model , although uses score produced training_set_1, still in unbiased way , since later never trained on these labels!

you might find paper quite useful