The Prediction of Merged Attributes with Multiple Viewpoint Systems

Multiple viewpoint systems find statistical structure in multi-dimensional entities, such as music, by combining Markov-based models together in order to make probabilistic predictions. This paper empirically tests two contrasting techniques for predicting multiple attributes of a musical surface. The first, an established method, predicts each attribute in turn, whilst a second, a proposed alternative, merges attributes into a new representation in order to make predictions simultaneously. A set of optimal smoothing techniques are found for both prediction methods across several harmonic and melodic datasets. Results indicate that when surface attributes are highly correlated, predicting merged attributes outperforms predicting the attributes separately. This can allow viewpoint systems with correlated surface attributes to be optimized, giving a closer fit with the training data as measured by mean information content.