Different Results With the Same Experiment about liblinear-java

[bwaldvogel]

Can you upload the datasets you are using?
Depending on your platform and compiler, I wouldn’t exclude a slightly different floating point precision/behaviour as an explanation.

[hrzafer]

Here is the link A bias term f=1 is added as the last feature.

[bwaldvogel]

Assuming you are using the "L1-regularized logistic regression" solver ("-s 6") I’ve conducted the following experiment with your dataset:

$ ./gradlew assemble
$ java -cp build/classes/main de.bwaldvogel.liblinear.Train -s 6 -p 0.3 train.libsvm model

iter   1  #CD cycles 1
iter   2  #CD cycles 1
iter   3  #CD cycles 1
iter   4  #CD cycles 1
iter   5  #CD cycles 3
iter   6  #CD cycles 1
=========================
optimization finished, #iter = 6
Objective value = 14871.7
#nonzeros/#features = 6455/181257

Whenever I repeat it, I get the exact same output and model file.
What are you doing differently?
Note: When you are training programmatically, you need to call Linear.resetRandom() if you want to get completely deterministic results.

[hrzafer]

I had also suspected that some kind of resetting mechanism was missing. I didn't know about the Linear.resetRandom() method which resolved the problem. Thanks.