This work investigates a new approach for symbolic music classification via continuous Haar wavelet transform, used as alternative representation and segmentation technique. The method was experimentally tested on the J.S. Bach’s 15 two-part inventions for keyboard. In music theory, the inventions are often described as developed coherently from a short melodic pattern that dominates the whole piece. Based on this premise, we take fragments of the inventions and try to identify the invention they belong to using a nearest neighbour classifier, which is based on the initial sections of all inventions. We compare combinations of wavelet and pitch vector representation and segmentations, as well as different lengths for the initial sections, and different ways of segmenting the sections and representing the segments, including melodic variations. The classification rates achieved by the combination of wavelet representation and segmentation are far better than those of pitch vectors and constant length segmentation, while including melodic variations yields only minimal rate changes.