lijin118 / lctc Goto Github PK

I'm wondering if the authors would like to release the code and data for the baseline method SVMt. I've tried using libsvm with linear kernel and other default parameters to reproduce the results in the paper but only got results as follows (mean and std over 10 random trials):
Caltech: 24.06 (6.4); Amazon: 37.26 (5.23); W: 54.72 (5.68). These results are worse than those reported in the paper (Caltech: 30.1 (0.9); Amazon: 44.2(1.1); W: 58.3(1.2)) in terms of both mean and std. I'd like to know what leads to such difference. Any comments will be much appreciated.

Hello，I directly use the demo to train the model:

model = train(full(Yref),sparse(St(:,1:size(Xref,1))'),sprintf('-s 0 -c %f -q 1',100));
[~,acc,~] = predict(full(Yt),sparse(St(:,size(Xref,1)+1:end)'),model);
fprintf('accuracy=%0.4f \n\n',acc);
The accuracy is:
start training, random runs 2
Accuracy = 20.7547% (55/265)
I change liblinear to libsvm and the accuracy is:
start training, random runs 3
svm accuracy=34.3396

However, when I want to utilize source domain data Ss to improve the performance:
model = train(full(Yref),sparse(St(:,1:size(Xref,1))'),sprintf('-s 0 -c %f -q 1',100));
model = train([full(Ys);full(Yref)],sparse([Ss';St(:,1:size(Xref,1))']),sprintf('-s 0 -c %f -q 1',100));
fprintf('accuracy=%0.4f \n\n',acc);
The accuracy is:
Accuracy = 7.92453% (21/265)
for libsvm, the accuracy is:
accuracy=15.0943

As the demo does not show how to use source domain data for training, I cannot obtain the experimental results in the paper. Could you tell me how to use the Data Ss for training? Thank you very much.