I am going to use LIBSVM for classification of data that uses a 'feature window'. Currently my data is in the form of a M x N x 3 matrix - M instances, N features and for each feature I'm considering the previous, current and next instance. If I were to only consider one instance per feature, I would simply feed the algorithm a list of instances described by N values, but in my case, each value is a 3-tuple.
Should I simply flatten all feature windows and use a M x 3N matrix? Would that have any impact on the learning process?