I was trying to build a classifier for a set of documents using a support vector machine. I choose to build the feature space using term occurrence. While experimenting, I found the following scenario: When removing stop words, the svm-based classifier was successfully built; otherwise, when keeping stop words, the SVM just could not be built and I got an error message “no support vector can be found”. I am very confused about this scenario. What might be the possible reason for this scenario?
It sounds like you might be doing "hard-margin" SVM. In this paradigm, I believe a support vector will only be found if perfect linearly separating hyperplane can be found in the data. Maybe try doing soft-margin SVM, which allows for errors (keep in mind that this approach adds a cost parameter you'll probably have to optimize in cross-validation, if you don't just use default settings).