An approach to classifying four-part music in multidimensional space - Emerging Trends in Image Processing, Computer Vision, and Pattern Recognition - page 29

Bach Versus Dykes Versus Mason

Technique 19 Metrics (%) 5 Metrics (%) Δ

Unweighted Points 69 67.1 − 1.9

Weighted Points 61.8 73.2 + 11.4

Euclidean Distance 50.5 70.4 + 19.9

Bach Versus Smart Versus Mason

Technique 19 Metrics (%) 5 Metrics (%) Δ

Unweighted Points 71.9 67.5 − 4.4

Weighted Points 70.4 71.9 + 1.5

Euclidean Distance 53.5 66.3 + 12.8

Dykes Versus Smart Versus Mason

Technique 19 Metrics (%) 5 Metrics (%) Δ

Unweighted Points 55.9 58 + 2.1

Weighted Points 57.6 57.9 + 0.3

Euclidean Distance 53.6 57.6 + 4

Bach Versus Dykes Versus Smart Versus Mason

Technique 19 Metrics (%) 5 Metrics (%) Δ

Unweighted Points 54.5 55.1 + 0.6

Weighted Points 52.6 57.6 + 5

Euclidean Distance 46.3 58.4 + 12.1

As seen above, the results with Mason were in keeping with the correctness levels of previ-
ous classiications, and as a whole, the ive-metric subset continued to perform beter than the
entire set of metrics. Not surprisingly, the percent of correct answers (for all techniques and
both metric sets) was overall at its lowest in the last set of trials, when all four composers were
classiied together. However, the percentages were well above 25% and statistically appropri-
ate for the change from three to four composers based on past three-way classifications. They
were actually much beter than expected.
A look through the classifier data produced during these seven classifications revealed
that Mason had a very high ThirdAppearsOnlyOnceInSAT percentage compared to the other
composers, yet a relatively low standard deviation. This raised the question of whether the
ThirdAppearsOnlyOnceInSAT metric was worth adding to the special set of five metrics, so