Options
"Test SVM with different attributes that train? Mistake??"
Maven
Hello,
I have trained an SVM (regression) with 28 attributes, optimized the values of C and G and see some nice performance.
My data file has 33 attributes and I use the RM process to remove 5 of them so that I train on the 28.
When testing with real world data, I made a MISTAKE and left 29 attributes in the test data. The resulting performance on real world tests was significantly better!!! (I tried removing the mistaken attribute from the test data, and performance
dropped.)
So, I tried to train the original model with the 29 attributes instead. The results were not as good.
The best performing model seems to be train with 28 attributes and test with 29. This seems strange.
How is this possible? I would assume that the SVM model applier in RM would have simply ignored any attributes not part of the saved model. (I'm very happy with the current results, but would like to understand what is happening.)
Can anyone explain this?
I have trained an SVM (regression) with 28 attributes, optimized the values of C and G and see some nice performance.
My data file has 33 attributes and I use the RM process to remove 5 of them so that I train on the 28.
When testing with real world data, I made a MISTAKE and left 29 attributes in the test data. The resulting performance on real world tests was significantly better!!! (I tried removing the mistaken attribute from the test data, and performance
dropped.)
So, I tried to train the original model with the 29 attributes instead. The results were not as good.
The best performing model seems to be train with 28 attributes and test with 29. This seems strange.
How is this possible? I would assume that the SVM model applier in RM would have simply ignored any attributes not part of the saved model. (I'm very happy with the current results, but would like to understand what is happening.)
Can anyone explain this?
Tagged:
0
