I've run into a curious situation that I wanted to ask
about. In using j48.J48, we have noticed that the order
of the attributes in the arff file seems to affect the
resulting tree that is learned. In other words, if you
reverse the order of the attributes in the arff file (and
adjust the data accordingly) a different tree is learned
that was learned with the original ordering.
The only possible explanation I have come up with is
that perhaps a "tie-breaking" mechanism in J48 is
based on the order in which attributes are defined. In
other words, if two attributes are considered equally
good at partitioning the training data at a particular
point in the learning process, then is the one that is
selected for inclusion in the tree based on its "position"
in the arff file? We observe this behavior consistently
so we don't think the "tie-breaking" is random.
Thanks for any tips or hints. I am just guessing on the
"tie-breaking" idea, so if I'm way off please say so! I have
looked at the Weka book and Quinlan's C4.5 book and didn't
see anything that addressed this particular issue (although
I will admit to having looked rather quickly).
Get your FREE download of MSN Explorer at http://explorer.msn.com/intl.asp
I'm new in Weka, and after 2 days of searching I mail to the big Weka-world and hope I will find the answers...
My primary problem is:
I make a standard neural net in the Kowledge Explorer, and I have an acceptable error rate.
Now I make the same neural net from the Command line - same parameters, same data - and the error rate is horrible. Mustn't there be the same error rate than in the Konwledge Explorer?
The only reason that I use the Command line is that I've not found how to generate Classifikation output Files with the Knowledge Explorer. I've only found the -d parameter with the command line.
My string looks as follows:
java -cp c:\Programme\weka-3-2\weka.jar weka.classifiers.neural.NeuralNetwork -L 0.3 -M 0.2 -N 500 -V 0 -S 0 -E 20 -H a -t AuswertungDrum.arff -d Klassifikator
makes same as follows:
java -cp c:\Programme\weka-3-2\weka.jar weka.classifiers.neural.NeuralNetwork -t AuswertungDrum.arff -d Klassifikator
The best solution for me would be the edition of the neural network into a file directly from the Knowledge Explorer out.
I've found under "Classify->more options" the checkbox "output model". But where is the output?
Thank you very much for your answers,
1.000.000 DM gewinnen - kostenlos tippen - http://millionenklick.web.de
IhrName(a)web.de, 8MB Speicher, Verschluesselung - http://freemail.web.de