Published by Springer, 2010. Hardcover (194 pages). ISBN: 978-3-642-13286-5.
Book Foreword
In the last 15 years we have seen a major transformation in the world of music Musicians use inexpensive personal computers instead of expensive recording studios to record, mix and engineer music. Musicians use the Internet to distribute their music for free instead of spending large amounts of money creating CDs, hiring trucks and shipping them to hundreds of record stores. As the cost to create and distribute recorded music has dropped, the amount of available music has grown dramatically. Twenty years ago a typical record store would have music by less than ten thousand artists, while today online music stores have music catalogs by nearly a million artists.
While the amount of new music has grown, some of the traditional ways of finding music have diminished. Thirty years ago, the local radio DJ was a music tastemaker, finding new and interesting music for the local radio audience. Now radio shows are programmed by large corporations that create playlists drawn from a limited pool of tracks Similarly, record stores have been replaced by big box retailers that have ever-shrinking music departments. In the past, you could always ask the owner of the record store for music recommendations. You would learn what was new, what was good and what was selling. Now, however, you can no longer expect that the teenager behind the cash register will be an expert in new music, or even be someone who listens to music at all.
With so much more music available, listeners are increasingly relying on tools such as automatic music recommenders to help them find music. Instead of relying on DJs, record store clerks or their friends to get music recommendations, listeners are also turning to machines to guide them to new music. This raises a number of questions:
- How well do these recommenders work?
- Do they generate novel, interesting and relevant music recommendations?
- How far into the Long Tail do they reach?
- Do they create feedback loops that drive listeners to a diminishing pool of popular artists?
- What affect will automatic music recommenders have on the collective music taste?
In this book, Dr Celma guides us through the world of automatic music recommendation. He describes how music recommenders work, explores some of the limitations seen in current recommenders, offers techniques for evaluating the effectiveness of music recommendations and demonstrates how to build effective recommenders by offering two real-world recommender examples. As we rely more and more on automatic music recommendation it is important for us to understand what makes a good music recommender and how a recommender can affect the world of music. With this knowledge we can build systems that offer novel, relevant and interesting music recommendations drawn from the entire world of available music.
Paul Lamere,
Director of Developer Community, The Echo Nest
Austin, TX, March 2010
Table of Contents
1. Introduction Page 1
1.1 Motivation 1
1.1.1 Academia 2
1.1.2 Industry 3
1.2 What's the problem with music recommendation? 4
1.3 Our proposal 6
1.4 Summary of contributions 8
1.5 Book outline 10
2. The recommendation problem 15
2.1 Formalisation of the recommendation problem 15
2.2 Use cases 16
2.3 General model 17
2.4 User profile representation 17
2.4.1 Initial generation 18
2.4.2 Maintenance 21
2.4.3 Adaptation 22
2.5 Recommendation methods 22
2.5.1 Demographic filtering 22
2.5.2 Collaborative filtering 23
2.5.3 Content-based filtering 28
2.5.4 Context-based filtering 30
2.5.5 Hybrid methods 34
2.6 Factors affecting the recommendation problem 35
2.7 Summary 38
3. Music recommendation 43
3.1 Use Cases 43
3.1.1 Artist recommendation 44
3.1.2 Playlist generation 44
3.1.3 Neighbour recommendation 45
3.2 User profile representation 45
3.2.1 Type of listeners 46
3.2.2 Related work 47
3.2.3 User profile representation proposals 48
3.3 Item profile representation 52
3.3.1 The Music Information Plane 53
3.3.2 Editorial metadata 55
3.3.3 Cultural metadata 56
3.3.4 Acoustic metadata 63
3.4 Recommendation methods 69
3.4.1 Collaborative filtering 70
3.4.2 Context-based filtering 73
3.4.3 Content-based filtering 75
3.4.4 Hybrid methods 78
3.5 Summary 80
4. The Long Tail in recommender systems 87
4.1 Introduction 87
4.2 The Music Long Tail 88
4.3 Definitions 93
4.3.1 Qualitative, informal definition 94
4.3.2 Quantitative, formal definition 95
4.4 Characterising a Long Tail distribution 97
4.5 The dynamics of the Long Tail 100
4.6 Novelty, familiarity and relevance 101
4.6.1 Recommending the unknown 102
4.6.2 Related work 104
4.7 Summary 105
5. Evaluation metrics 109
5.1 Evaluation strategies 109
5.2 System-centric evaluation 110
5.2.1 Predictive-based metrics 110
5.2.2 Decision-based metrics 111
5.2.3 Rank-based metrics 113
5.2.4 Limitations 115
5.3 Network-centric evaluation 116
5.3.1 Navigation 117
5.3.2 Connectivity 118
5.3.3 Clustering 120
5.3.4 Centrality 121
5.3.5 Limitations 122
5.3.6 Related work in Music Information Retrieval 123
5.4 User-centric evaluation 123
5.4.1 Gathering feedback 124
5.4.2 Limitations 125
5.5 Summary 126
6. Network-centric evaluation 129
6.1 Network analysis and the Long Tail model 129
6.2 Artist network analysis 131
6.2.1 Datasets 131
6.2.2 Network analysis 132
6.2.3 Popularity analysis 139
6.2.4 Discussion 145
6.3 User network analysis 146
6.3.1 Datasets 146
6.3.2 Network analysis 148
6.3.3 Popularity analysis 151
6.3.4 Discussion 154
6.4 Summary 155
7. User-centric evaluation 157
7.1 Music Recommendation Survey 157
7.1.1 Procedure 157
7.1.2 Datasets 158
7.1.3 Participants 159
7.2 Results 160
7.2.1 Demographic data 160
7.2.2 Quality of the recommendations 161
7.3 Discussion 165
7.4 Limitations 166
8. Applications 169
8.1 Searchsounds: Music discovery in the Long Tail 169
8.1.1 Motivation 169
8.1.2 Goals 171
8.1.3 System overview 172
8.1.4 Summary 175
8.2 FOAFing the Music: Music recommendation in the Long Tail 175
8.2.1 Motivation 175
8.2.2 Goals 176
8.2.3 System overview 177
8.2.4 Summary 182
9. Conclusions and Further Research 185
9.1 Book Summary 186
9.1.1 Scientific contributions 186
9.1.2 Industrial contributions 188
9.2 Limitations and Further Research 189
9.3 Outlook 191