Learning Similarity Measures for Music

Musical similarity is a central issue in Music Information Retrieval
(MIR) and the key to many applications. In the classical retrieval
scenario, similarity is used as an estimate for relevance to rank a list
of songs or melodies. Further applications comprise the sorting and
organization of music collections by grouping similar music pieces or
generating maps for collection overview. Finally, music recommender
systems that follow the popular “find me more like…”-idea often employ
a similarity-based strategy as well.
However, music similarity is not a simple concept. For a comparison of
music pieces, many interrelated features and facets can be considered.
Their individual importance and how they should be combined depends very
much on the user and his specific retrieval task. Users of MIR systems
may have a varying (musical) background and experience music in
different ways. Consequently, when comparing musical pieces with each
other, opinions may diverge. A musician, for instance, might especially
look after structures, harmonics or instrumentation (possibly paying –
conscious- or unconsciously – special attention to his own instrument).
Non-musicians will perhaps focus more on overall timbre or general mood.
Others, in turn, may have a high interest in the lyrics as long as they
are able to understand the particular language. Apart from considering
individual users, similarity measures also should be tailored for their
specific retrieval task to improve the performance of the retrieval
system. For instance, when looking for cover versions of a song, the
timbre may be less interesting than the lyrics.
In order to support individual user perspectives and different retrieval
tasks, music similarity should no longer be considered as a static
element of MIR systems. Therefore, this tutorial aims to introduce MIR
researchers and developers to a variety of existing techniques, which
allow them to build systems including an adaptable model of music
similarity. We show how to model the task of finding a suitable music
similarity measure as a machine learning problem, introduce various
learning algorithms, and give examples of interactive applications.
The tutorial addresses a broad audience and in particular does not
require prior knowledge in machine learning as a prerequisite. The ideas
and algorithms covered will be presented up to the level of
understanding that enables the audience to successfully apply them.
Proposed outline for a full-day tutorial:
1. Introduction and Basics
1.1. Definition / Concepts of Similarity
1.2. Features and Facets of Music
1.3. Quantifying Similarity
1.4. Designing Similarity Studies
2. Learning Linear Similarity Models (Sebastian)
3. Learning Mahalanobis Distances (Daniel)
4. Designing Application and User Interfaces with Adaptive Music
Similarity Measures
+ Optional advanced topics (e.g., feedback based training methods,
active learning, and learning local models)