Unsupervised change detection in data streams: an application in music analysis

The mining of data streams has been attracting much attention in the recent years, specially from Machine Learning researchers. One important task in learning from data streams is to correctly detect changing data characteristics over time, since this is critical to the correct modeling of data behavior. With the understanding that many applications generate unlabeled streams, different algorithms have been proposed to approach unsupervised change detection. These algorithms implement different strategies, from simple incremental methods that monitor data statistics, to more advanced techniques based on divergences of clustering models. In recent studies, however, authors pointed out those algorithms lack in learning guarantees, meaning that results obtained by these methods could be due to model parameterization. These observations led to the development of a new stability concept that is suitable for unsupervised streams. This stability concept motivated a new change detection algorithm which ensures model modifications corresponding to actual data changes. Previous results on artificial scenarios have confirmed this algorithm’s ability to correctly detect changes. However, the requirement of assessing the algorithm’s performance on real-world data remained, which is essential to the understanding of the algorithm’s capabilities. Motivated by this observation, this work applied this algorithm to the domain of audio analysis, more specifically, in music change detection. Results obtained in different music tracks provide interesting insights on the types of changes that produce a more significant impact on the algorithm’s decisions, allowing for a better understanding about its underlying dynamics.