No need to test for positive and negative separately. One class SVM can be trained by using just one class and it has the ability to classify a data point into either positive or negative regardless of how it is tested. If you are getting more +1 then it may be probably because of underfitting. Did you try by optimizing the SVM parameters? If not, you may try tuning the parameters as they critically affect the nature of fitting. 

Sure, I optimized parameter gamma using cross validation but I don't know how to set the exactly parameter nu. I choose nu by 0.01 or sometimes 0.05.

as your opinion the svm-scaling is important ?

Dear Ehsan, 

As you know one-class SVM is mainly used for anomaly detection and is applied when we have a large population of normal cases but only few outliers. You do not need to train the model separately and can use the whole data. If you fine tune the nu parameter  imperatively, it should correctly reveal the largest number of anomalies with test data. You can use cross validation to find the right parameter.

I also suggesy having a look at the following link:

https://msdn.microsoft.com/en-us/library/azure/dn913103.aspx

Good luck!

By setting nu to 0.01 guarantees to find at most 1% of the training data being misclassified. In other word nu sets the upper bound on the fraction of margin errors. Therefor we can force the fitting by lowering the nu value, which also can be result in overfitting. With a higher value like 0.05, we expect at most 5% of the training data can be misclassified and can be result in underfitting. If the size of training data is very large (as it well represents the positive space) selecting a smaller nu will be better. large nu value can be used with weak training datasets.