aset400 - heterogeneous embedding.tgz

Format: 

Matlab

Number of points: 

412

Number of kernels: 

5

Tasks: 

Binary

Performance: 

0.79

Measure: 

Accuracy

Best paper: 

Heterogeneous Embedding for Subjective Artist Similarity

References: 

Heterogeneous Embedding for Subjective Artist Similarity

The Quest for Ground Truth in Musical Artist Similarity

Semantic Annotation and Retrieval of Music and Sound Effects

Partial order embedding with multiple kernels

This data set contains the aset400 artists and subjective similarity measurements from Ellis, et al 1.  Also included are 5 kernels derived from text and acoustic features associated with the artists, as described in 2. The variables in the .mat file are as follows:

• Kbio and Ktag are TF-IDF kernels computed from stemmed Last.FM biographies and top-100 tags. 
  • bag_* and dict_* contain the bag-of-words and dictionaries for each type of document. 
  • T* contain the TF-IDF filtered bags-of-words.
• Kmfcc is a probability product kernel (PPK) over the Gaussian mixture models (contained in models).
• Ksm is a PPK derived from semantic multinomials (SM).
• Kchroma is a KL-divergence pseudo-kernel derived from chroma features.
• names is a dictionary of the names of the artists in question.

Similarity measurements are derived from aset400, and encoded as follows:

• indices.csv contains the list of artist index numbers
• triples.csv contains the similarity triples, one per line, encoded by index number.  A line "X,Y,Z" indicates that the similarity between (X,Y) is greater than between (X,Z).

• 1. [18] Ellis, D. P. W., Whitman B., Berenzweig A., & Lawrence S. (2002).  The Quest for Ground Truth in Musical Artist Similarity. ISMIR 2002, 3rd International Conference on Music Information Retrieval.
• 2. [17] McFee, B., & Lanckriet G. (2009).  Heterogeneous Embedding for Subjective Artist Similarity. Tenth International Symposium for Music Information Retrieval (ISMIR).