Supporting Folk-Song Research by Automatic Metric Learning and Ranking

Korinna Bade, Andreas Nürnberger, Sebastian Stober, Jörg Garbers, Frans WieringISMIR 2009

Introduction

Folk song research, particularly the study of tune evolution through oral transmission, relies heavily on the classification of songs into "melody norms"—families of related variants. While computational musicology offers various similarity measures (e.g., edit distance, rhythmic patterns), no single algorithm captures the nuanced, subjective criteria used by expert musicologists. This research addresses the gap by proposing a semi-supervised metric learning approach that optimizes a "super-metric" composed of multiple underlying musical features.

Methodology: Metric Learning via MLB Constraints

The core of the proposed system is a weighted linear combination of 14 distinct similarity measures. Rather than selecting a single "best" metric, the system learns an optimal weight vector ($w$) to combine them.

The weights are trained using Must-Link-Before (MLB) constraints derived from expert classifications. For a given query song $q$, any song $t_r$ belonging to the same melody norm must be ranked as more similar than any song $t_i$ from a different class. A gradient descent algorithm minimizes violations of these expert-derived rules.

Specificity Levels

The authors evaluate three levels of weight adaptation:

  1. Query-specific ($w^q$): Weights optimized for an individual song.
  2. Class-specific ($w^{cl}$): Weights optimized for an entire melody norm family.
  3. Overall ($w^a$): A global weighting scheme for the entire dataset.

For unclassified songs, the system employs a case-based approach, identifying the nearest known neighbor in the database and adopting its optimized weighting scheme.

Analysis of Musical Similarity

A key finding of this study is that musical similarity is highly localized. The features that define one melody norm (e.g., rhythmic structure in dance tunes) may be irrelevant for another (e.g., melodic contour in ballads).

Inner-class vs. cross-class similarity across 26 melody norms.

As shown in the figure above, the average pairwise inner-class similarity consistently exceeds cross-class similarity across the 26 melody norms studied. This disparity confirms that different song families require distinct mathematical "lenses" for accurate categorization, justifying the use of localized weighting schemes.

Experimental Results

The system was tested on a dataset of 360 melodic strophes pre-classified into 26 melody norms. The learned metrics consistently outperformed individual baseline measures like raw edit distance (rawEd).

Performance metrics for various weighting schemes across different test scenarios.

The results in the table indicate that query-specific ($w^q$) and class-specific ($w^{cl}$) weights provide the highest precision, significantly reducing the number of "1st Wrong" classifications compared to the baseline rawEd. Even the "overall" weight ($w^a$) provides a measurable improvement, demonstrating that the learning algorithm captures genuine musical characteristics rather than just memorizing the training data.

Conclusion and Future Directions

The study demonstrates that a learned combination of metrics better reflects expert musicological intuition than any single algorithm. While query-specific weighting offers an upper bound on performance, class-specific weights provide a practical and effective middle ground. Future work will focus on using more granular musical "atoms" as building blocks for the learning algorithm and exploring how weight clusters might reveal undocumented sub-melody norms.

Summary updated

BibTeX
@inproceedings{bade2009supporting,
  author = {Korinna Bade and Andreas Nürnberger and Sebastian Stober and Jörg Garbers and Frans Wiering},
  title = {{Supporting Folk-Song Research by Automatic Metric Learning and Ranking}},
  booktitle = {ISMIR},
  year = {2009},
}
Figure