Articles - cscenter/automatic-playlist-generation GitHub Wiki

A Comparative Study on Content-Based Music Genre Classification (Tao Li, Mitsunori Ogihara, Qi Li)

Overview

Features used

Learning method

Playlist generation algorithm

Relevance to our needs

Useful references

Automatic Generation of Social Tags for Music Recommendation (Douglas Eck, Paul Lamere, Thierry Bertin-Mahieux, Stephen Green)

Overview

Features used

Learning method

Playlist generation algorithm

Relevance to our needs

Useful references

Evaluating the Quality of Playlists Based on Hand-Crafted Samples (Geoffray Bonnin, Dietmar Jannach)

Overview

Features used

Learning method

Playlist generation algorithm

Relevance to our needs

Useful references

Generating Music Playlists with Hierarchical Clustering and Q-Learning (James King, Vaiva Imbrasaite)

Overview

Features used

Learning method

Playlist generation algorithm

Relevance to our needs

Useful references

Learning a Gaussian Process Prior for Automatically GeneratingMusic Playlists (John C. Platt, Christopher J. C. Burges, Steven Swenson, Christopher Weare, Alice Zheng)

Overview

Authors introduced the new method of generating a kernel for use on Gaussian Process Regression adding the meta-training phase.

Features used

  • Genre, 1 value (e.g. Jazz, Reggae, Hip-Hop)
  • Subgenre, 1 value (e.g. Heavy Metal, I’m So Sad and Spaced Out)
  • Style, 1 value (e.g. East Coast Rap, Gangsta Rap, West Coast Rap)
  • Mood, 1 value (e.g. Dreamy, Fun, Angry)
  • Rhythm Type, 1 value (e.g. Straight, Swing, Disco)
  • Rhythm Description, 1 value (e.g. Frenetic, Funky, Lazy)
  • Vocal Code, 1 value (e.g. Instrumental,Male, Female, Duet)

Learning method

Meta-learning phase is being done over a large set of existing playlists. Similarity of song in kernel is being measured by equality of features

Playlist generation algorithm

Kernel is being learned on large set of predefined playlists. First step is to compute the Gaussian statistics over a set of functions related to one's target. Secondly, one compute the estimation for the kernel parameters After learning the kernel and achieving the seed vector one computes the hyperparameter for noise, and after that the best-of-n tracks represents the playlist

Relevance to our needs

We are not using the existing playlists for learning, and features were assigned by the editors.

Useful references

  • T. Minka and R. Picard - Learning how to learn is learning with points sets (fitting the Gaussian Process)

Learning a Music Similarity measure on Automatic Annotations with Application to Playlist Generation (Linxing Xiao, Lie Lu, Frank Seide, Jie Zhou)

Overview

Authors introduced the learning similarity measuring technique with kernel-based similarity being optimized with Non-Negative Quadratic Optimization (which is mainly based on Gradient Projection)

Features used

Authors used annotation techniques from their previous work (Collective annotation of music from multiple semantic categories):

  • Genre, 1-2 values from [Blues, Country, Electronica, Folk, Funk, Gospel, HardRock, Jazz, Pop, Punk, Rap, R&B, Rock-roll, SoftRock]
  • Instrument, 1-5 values from [Acoustic Guitar, Acoustic Piano, Bass, Drum, Electric Guitar, Electric Piano, Harmonica, Horn, Organ, Percussion, Sax, String]
  • Texture, 1-2 values from [Acoustic, Electric, Synthetic]
  • Vocal, 1-2 values from [Group, Male, Female, None]
  • Affective, 1 value from [Positive, Neutral, Negative]
  • Arousal, 1 value from [Strong, Middle, Weak]
  • Rhythm, 1 value from [Strong, Middle, Weak]
  • Tempo, 1 value from [Fast, Moderato, Slow]
  • Tonality, 1 value from [Major, Mixed, Minor]
  • Production, 1 value from [Studio, Live]

Learning method

Kernel based similarity measure

Playlist generation algorithm

Rank each track with weighted sum of similarities

Relevance to our needs

Should use their previous work for feature extraction

Useful references

  • Z. Y. Duan, L. Lu, and C. S Zhang - "Collective annotation of music from multiple semantic categories" (feature extraction)
  • K. Kaji, K. Hirata, and Nagao K. - "A music recommendation system based on annotations about listeners’ preferences and situations" (cousine measure as base line)
  • E. Pampalk, S Dixon, and G. Widmer - "On the evaluation of perceptual similarity measures for music" (low-level song features)
  • Other articles from references section regarding the mid-level music attributes

Learning to Embed Songs and Tags for Playlist Prediction (Shuo Chen, Joshua L. Moore, Thorsten Joachims, Douglas Turnbull)

Overview

Features used

Learning method

Playlist generation algorithm

Relevance to our needs

Useful references

Playlist Prediction via Metric Embedding (Shuo Chen, Joshua L. Moore, Thorsten Joachims, Douglas Turnbull)

Overview

Features used

Learning method

Playlist generation algorithm

Relevance to our needs

Useful references

Steerable Playlist Generation by Learning Song Similarity From Radio Station Playlists (Franc¸ois Maillet, Douglas Eck, Guillaume Desjardins, Paul Lamere)

Overview

Features used

Learning method

Playlist generation algorithm

Relevance to our needs

Useful references