Multi-dimensional Banded Pattern Mining

Techniques for identifying "banded patterns" in n-Dimensional (n-D) zero-one data, so called Banded Pattern Mining (BPM), are considered. Previous work directed at BPM has been in the context of 2-D data sets; the algorithms typically operated by considering permutations which meant that extension to n-D could not be easily realised. In the work presented in this paper banding is directed at the n-D context. Instead of considering large numbers of permutations the novel approach advocated in this paper is to determine banding scores associated with individual indexes in individual dimensions which can then be used to rearrange the indexes to achieve a "best" banding. Two variations of this approach are considered, an approximate approach (which provides for efficiency gains) and an exact approach.