Representation of uncertain and imprecise behaviors and its application to music performance

Abstract

Many phenomena in the real world can be understood as behaviors, because they follow some underlying rules and behave in characteristic ways. The computational representation and classification of them is a task of great interest for researchers in behavioral sciences as they are in continuous need of new tools and methods to represent and understand behaviors of growing complexity. With mobile devices becoming ubiquitous, data series are gaining in importance as a suitable representation for the information coming from their sensors. Data series are defined as an ordered sequence of data at given intervals of an indexing variable (e.g. time). Behaviors can be defined as imprecise and uncertain multivariate data series. Although the problem of representing imperfect data has been addressed many times in the past, the lack of a general and universal solution obligates to build ad-hoc systems for different problems. For that reason, there is still a need to develop new models to represent such information. Our proposal, which assumes that some kind of commonality exists among instances of the same behavior in a given domain, represents those imperfect data series as a set of probability distributions. To do so, it first transforms the imperfect observations into qualitative values. Then, it selects a dimension of the behavior and uses it to look for correlations with the rest of dimensions. These correlations are expressed as discrete probability distributions. The model aims to be general enough to be employed in any domain that contains imprecise and uncertain behaviors, but we concentrated in the particular domain of music in order to validate it. Experiments showed that our representation allows to identify violinists in a dataset of monophonic violin recordings from 23 well-known performers, outperforming comparable alternatives. Departing from that application, we also studied several aspects of the computational representation of music performances, identified the chorus of songs by means of mining frequent patterns, and proposed a framework for automatic music composition.