In Search of the Consensus Among Musical Pattern Discovery Algorithms

Patterns are an essential part of music and there are many different algorithms that aim to discover them. Based on the improvements brought by using data fusion methods to find the consensus of algorithms on other MIR tasks, we hypothesize that fusing the output from musical pattern discovery algorithms will improve the pattern discovery results. In this paper, we explore two methods to combine the pattern output from ten state-of-the-art algorithms using two datasets. Both provide human-annotated patterns as ground truth. We show that finding the consensus among the output of different musical pattern discovery algorithms is challenging for two reasons: First, the number of patterns found by the algorithms exceeds patterns in human annotations by several orders of magnitude, with little agreement on what constitutes a pattern. Second, the algorithms perform inconsistently across different pieces. We show that algorithms lack a consensus with each other. Therefore, it is difficult to harness the collective wisdom of the algorithms to find ground truth patterns. The main contribution of this paper is a meta-analysis of the (dis)similarities among pattern discovery algorithms’ output and using the output in two fusion methods. Furthermore, we discuss the implication of our results for the MIREX task.