The output of the "predict" function of an SVM implementation will be a vector of size N, where N is the number of examples you are trying to get predictions for. The numbers in this vector correspond to each example's distance to the class-separating hyperplane. In order to dichotomize these values so that each is a binary corresponding to class membership, you need to use a prediction threshold. Any number greater than or equal to that threshold will be predicted to be part of the positive class and any number less than that threshold will be predicted to be part of the negative class. The definition of the SVM algorithm will ensure that a prediction threshold of 0 is the optimal threshold for the training data (i.e., it yields the highest classification accuracy). However, for generalizing to independent data sets, you may want to adjust the prediction threshold and see how your performance metric of choice (e.g., accuracy, precision, or recall) is affected. I suggest writing a simple function that accepts data and a prediction threshold and outputs a performance metric. You can then try different thresholds by looping over a range from the minimum to the maximum output value:
linspace(min(output),max(output),10)
