Using Transportation Distances for Measuring Melodic Similarity Pichaya Tappayuthpijarn

Download Report

Transcript Using Transportation Distances for Measuring Melodic Similarity Pichaya Tappayuthpijarn

Using Transportation
Distances for Measuring
Melodic Similarity
Pichaya Tappayuthpijarn
Qiang Wang
Outline
Introduction
 Melodic Similarity Measuring
 The Earth Mover’s Distance (EMD)
 The Proportional Transportation Distance
(PTD)
 Results & Conclusion

Introduction
Weighted notes are well suited to
represent music.
 Measuring melodic similarity by comparing
weighted notes.
 Using EMD or PTD to measure the
similarity.

Melodic Similarity Measuring

Weighted Notes represented by (Time, Pitch,
Weight). The weights only reflect the note
durations.
Properties of a Metric
**For pseudo-metric, only properties i, iii and iv holds.
The Earth Mover’s Distance (EMD)

measures the minimum amount of work
needed to transform one into the other by
moving weight.
Minimize the overall cost
 d
iI jJ
with constraints
f
ij ij
W, U are the total weights of A, B
respectively,
Example : a weight flow with EMD
The Proportional Transportation
Distance (PTD)
EMD v.s. PTD
 PTD is a modified version of EMD.
 In the case of unequal total weight,
triangle inequality does not hold for the
EMD.

By PTD, triangle inequality holds, making
database search more efficient by using
indices.
The Proportional Transportation
Distance
for both point set, divide every point’s
weight by its point set’s total weight
 then calculate the EMD for the resulting
point sets.

Properties of PTD

A pseudo-metric.
Triangle inequality holds true
a+b>=c

Gives clear distinction of groups.
The fact that the EMD allows partial matching, while the
PTD matches all notes, leads to a clear distinction of
these groups by the PTD, but not the EMD.

Can recognize augmented versions of the
same melody as similar.
Result & Conclusion
PTD leads to the identification of about
3.9% of 100 randomly chosen search
( for r = 1).
 For r = 5, we need less than 1000
calculations instead of 476,000.


PTD Indexing reduces the query
running time.
Thank you!