Music Recommendation and Discovery in the Long Tail

14,793 views

Published on

Music consumption is biased towards a few popular artists. For instance, in 2007 only 1% of all digital tracks accounted for 80% of all sales. Similarly, 1,000 albums accounted for 50% of all album sales, and 80% of all albums sold were purchased less than 100 times. There is a need to assist people to filter, discover, personalise and recommend from the huge amount of music content available along the Long Tail.

Current music recommendation algorithms try to accurately predict what people demand to listen to. However, quite often these algorithms tend to recommend popular -or well-known to the user- music, decreasing the effectiveness of the recommendations. These approaches focus on improving the accuracy of the recommendations. That is, try to make accurate predictions about what a user could listen to, or buy next, independently of how useful to the user could be the provided recommendations.

In this Thesis we stress the importance of the user's perceived quality of the recommendations. We model the Long Tail curve of artist popularity to predict -potentially- interesting and unknown music, hidden in the tail of the popularity curve. Effective recommendation systems should promote novel and relevant material (non-obvious recommendations), taken primarily from the tail of a popularity distribution.

The main contributions of this Thesis are: (i) a novel network-based approach for recommender systems, based on the analysis of the item (or user) similarity graph, and the popularity of the items, (ii) a user-centric evaluation that measures the user's relevance and novelty of the recommendations, and (iii) two prototype systems that implement the ideas derived from the theoretical work. Our findings have significant implications for recommender systems that assist users to explore the Long Tail, digging for content they might like.

Published in: Technology, Education, Business
2 Comments
31 Likes
Statistics
Notes
No Downloads
Views
Total views
14,793
On SlideShare
0
From Embeds
0
Number of Embeds
158
Actions
Shares
0
Downloads
573
Comments
2
Likes
31
Embeds 0
No embeds

No notes for slide

Music Recommendation and Discovery in the Long Tail

  1. Music Recommendation and Discovery in the Long Tail Òscar Celma Doctoral Thesis Defense (Music Technology Group ~ Universitat Pompeu Fabra)
  2. PhD defense // UPF // Feb 16th 2009 Music Recommendation (personalized) and Discovery (explore large music collections) in the Long Tail (non-obvious, novel, relevant music)
  3. PhD defense // UPF // Feb 16th 2009 “The Paradox of Choice: Why More Is Less”, Barry Schwartz (2004) The problem Paradox of choice
  4. PhD defense // UPF // Feb 16th 2009 music overload • Today(August, 2007) iTunes: 6M tracks  P2P: 15B tracks  53% buy music on line  • Finding unknown, relevant music is hard! Awareness vs. access to content 
  5. PhD defense // UPF // Feb 16th 2009 music overload? Digital Tracks – Sales data for 2007 ● ● Nearly 1 billion sold in 2007 ● ● 1% of tracks account for 80% of sales ● ● 3.6 million tracks sold less than 100 copies, and ● 1 million tracks sold exactly 1 copy ● • • •Data from Nielsen Soundscan 'State of the (US) industry' 2007 report
  6. PhD defense // UPF // Feb 16th 2009 the Long Tail of popularity • Help me find it! [Anderson, 2006]
  7. PhD defense // UPF // Feb 16th 2009 research questions • 1) How can we evaluate/compare different music recommendation approaches? • 2) How far into the Long Tail do music recommenders reach? • 3) How do users perceive novel (unknown to them), non-obvious recommendations?
  8. PhD defense // UPF // Feb 16th 2009 If you like The Beatles you might like ...
  9. PhD defense // UPF // Feb 16th 2009
  10. PhD defense // UPF // Feb 16th 2009
  11. PhD defense // UPF // Feb 16th 2009
  12. PhD defense // UPF // Feb 16th 2009 • popularity bias • low novelty ratio
  13. PhD defense // UPF // Feb 16th 2009 FACTORS AFFECTING RECOMMENDATIONS: Novelty Relevance Diversity Cold start Coverage Explainability Temporal effects
  14. PhD defense // UPF // Feb 16th 2009 FACTORS AFFECTING RECOMMENDATIONS: Novelty Relevance Diversity Cold start Coverage Explainability Temporal effects
  15. PhD defense // UPF // Feb 16th 2009 novelty vs. relevance
  16. PhD defense // UPF // Feb 16th 2009 how can we measure novelty? • predictive accuracy vs. perceived quality • metrics MAE, RMSE, P/R/F-measure, ...  Test Train Can't measure novelty 
  17. PhD defense // UPF // Feb 16th 2009 how can we measure novelty? • predictive accuracy vs. perceived quality • metrics MAE, RMSE, P/R/F-measure, ...  Can measure novelty 
  18. PhD defense // UPF // Feb 16th 2009 how can we measure relevance? quot;The key utility measure is user happiness. It seems reasonable to assume that relevance of the results is the most important factor: blindingly fast, useless answers do not make a user happy.quot;  quot;Introduction to Information Retrievalquot; (Manning, Raghavan, and Schutze, 2008)
  19. PhD defense // UPF // Feb 16th 2009
  20. PhD defense // UPF // Feb 16th 2009 research in music recommendation • Google Scholar Papers that contain “music recommendation” or “music recommender” in the title (Accessed October 1st, 2008)
  21. PhD defense // UPF // Feb 16th 2009 research in music recommendation • ISMIR community
  22. PhD defense // UPF // Feb 16th 2009 music recommendation approaches • Expert-based • Collaborative filtering • Context-based • Content-based • Hybrid (combination)
  23. PhD defense // UPF // Feb 16th 2009 music recommendation approaches • Expert-based AllMusicGuide  Pandora  • Collaborative filtering • Context-based • Content-based • Hybrid (combination)
  24. PhD defense // UPF // Feb 16th 2009 music recommendation approaches • Expert-based • Collaborative filtering User-Item matrix  [Resnick, 1994], [Shardanand, 1995], [Sarwar, 2001] • Context-based • Content-based
  25. PhD defense // UPF // Feb 16th 2009 music recommendation approaches • Expert-based • Collaborative filtering User-Item matrix  [Resnick, 1994], [Shardanand, 1995], [Sarwar, 2001] Similarity   Cosine  Adj. cosine  Pearson  SVD / NMF: matrix factorization • Context-based • Content-based
  26. PhD defense // UPF // Feb 16th 2009 music recommendation approaches • Expert-based • Collaborative filtering User-Item matrix  [Resnick, 1994], [Shardanand, 1995], [Sarwar, 2001] Similarity   Cosine  Adj. cosine  Pearson  SVD / NMF: matrix factorization Prediction (user-based)   Avg. weighted
  27. PhD defense // UPF // Feb 16th 2009 music recommendation approaches • Expert-based • Collaborative filtering • Context-based WebMIR  thrash [Schedl, 2008] Content Reviews Lyrics Blogs heavy metal Tags Bios Playlists Social Edgy Weird concert 90s Loud rock [Hu&Downie, 2006] [Celma et al., 2006] [Levy&Sandler, 2007] [Baccigalupo, 2008] [Symeonidis, 2008] • Content-based • Hybrid (combination)
  28. PhD defense // UPF // Feb 16th 2009 music recommendation approaches • Expert-based • Collaborative filtering • Context-based • Content-based Audio features   Bag-of-frames (MFCC) [Aucouturier, 2004], Rhythm [Gouyon, 2005], Harmony [Gomez, 2006], ... Similarity   KL-divergence: GMM [Aucouturier, 2002]  EMD [Logan, 2001]  Euclidean: PCA [Cano, 2005]  Cosine: mean/var (feature vectors)  Ad-hoc • Hybrid (combination)
  29. PhD defense // UPF // Feb 16th 2009 music recommendation approaches • Expert-based • Collaborative filtering • Context-based • Content-based • Hybrid (combination) Weighted  Cascade  Switching 
  30. PhD defense // UPF // Feb 16th 2009 Work done
  31. PhD defense // UPF // Feb 16th 2009 contributions
  32. PhD defense // UPF // Feb 16th 2009 contributions 1) Network-based evaluation Item Popularity + Complex networks
  33. PhD defense // UPF // Feb 16th 2009 contributions 1) Network-based evaluation Item Popularity + Complex networks 2) User-based evaluation
  34. PhD defense // UPF // Feb 16th 2009 contributions 1) Network-based evaluation Item Popularity + Complex networks 2) User-based evaluation 3) Systems
  35. PhD defense // UPF // Feb 16th 2009 contributions
  36. PhD defense // UPF // Feb 16th 2009 contributions
  37. PhD defense // UPF // Feb 16th 2009 complex network analysis :: artists • 3 Artist similarity (directed) networks CF*: Social-based, incl. item-based CF (Last.fm)   “people who listen to X also listen to Y” CB: Content-based Audio similarity   “X and Y sound similar” EX: Human expert-based (AllMusicGuide)   “X similar to (or influenced by) Y”
  38. PhD defense // UPF // Feb 16th 2009 complex network analysis :: artists • 3 Artist similarity (directed) networks CF*: Social-based, incl. item-based CF (Last.fm)   “people who listen to X also listen to Y” CB: Content-based Audio similarity   “X and Y sound similar” EX: Human expert-based (AllMusicGuide)   “X similar to (or influenced by) Y”
  39. PhD defense // UPF // Feb 16th 2009 complex network analysis :: artists • Small-world networks [Watts & Strogatz, 1998] Network traverse in a few clicks 
  40. PhD defense // UPF // Feb 16th 2009 complex network analysis :: artists • Indegree – avg. neighbor indegree correlation r = Pearson correlation  [Newman, 2002]
  41. PhD defense // UPF // Feb 16th 2009 complex network analysis :: artists • Indegree – avg. neighbor indegree correlation
  42. PhD defense // UPF // Feb 16th 2009 complex network analysis :: artists • Indegree – avg. neighbor indegree correlation
  43. PhD defense // UPF // Feb 16th 2009 complex network analysis :: artists • Indegree – avg. neighbor indegree correlation Kin(Bruce Springsteen)=534 => avg(Kin(sim(Bruce Springsteen)))=463
  44. PhD defense // UPF // Feb 16th 2009 complex network analysis :: artists • Indegree – avg. neighbor indegree correlation Kin(Bruce Springsteen)=534 => avg(Kin(sim(Bruce Springsteen)))=463 Kin(Mike Shupp)=14 => avg(Kin(sim(Mike Shupp)))=15
  45. PhD defense // UPF // Feb 16th 2009 complex network analysis :: artists • Indegree – avg. neighbor indegree correlation Kin(Bruce Springsteen)=534 => avg(Kin(sim(Bruce Springsteen)))=463 Kin(Mike Shupp)=14 => avg(Kin(sim(Mike Shupp)))=15 Homophily effect!
  46. PhD defense // UPF // Feb 16th 2009 complex network analysis :: artists • Indegree – avg. neighbor indegree correlation Last.fm presents assortative mixing (homophily)   Artists with high indegree are connected together, and similarly for low indegree artists
  47. PhD defense // UPF // Feb 16th 2009 complex network analysis :: artists • Last.fm is a scale-free network [Barabasi, 2000] power law exponent for the cumulative indegree  distribution [Clauset, 2007] A few artists (hubs) control the network 
  48. PhD defense // UPF // Feb 16th 2009 complex network analysis :: artists • Summary: artist similarity networks |------------|---------|-----|-----------| | | Last.fm | CB | Exp (AMG) | |------------|---------|-----|-----------| |Small World | yes | yes | yes | | | | | | |Ass. mixing | yes | No | No | | | | | | | Scale-free | yes | No | No | |------------|---------|-----|-----------| Last.fm artist similarity network resembles to a social  network (e.g. facebook)
  49. PhD defense // UPF // Feb 16th 2009 complex network analysis :: artists • But, still some remaining questions... Are the hubs the most popular artists?  How can we navigate along the Long Tail, using  the artist similarity network?
  50. PhD defense // UPF // Feb 16th 2009 contributions Long Tail analysis
  51. PhD defense // UPF // Feb 16th 2009 the Long Tail in music • last.fm dataset (~260K artists)
  52. PhD defense // UPF // Feb 16th 2009 the Long Tail in music • last.fm dataset (~260K artists) the beatles (50,422,827) radiohead (40,762,895) red hot chili peppers (37,564,100) muse (30,548,064) death cab for cutie (29,335,085) pink floyd (28,081,366) coldplay (27,120,352) metallica (25,749,442)
  53. PhD defense // UPF // Feb 16th 2009 the Long Tail model [Kilkki, 2007] • F(x) = Cumulative distribution up to x
  54. PhD defense // UPF // Feb 16th 2009 the Long Tail model [Kilkki, 2007] • Top-8 artists: F(8)~ 3.5% of total plays 50,422,827 the beatles 40,762,895 radiohead 37,564,100 red hot chili peppers 30,548,064 muse 29,335,085 death cab for cutie 28,081,366 pink floyd 27,120,352 coldplay 25,749,442 metallica
  55. PhD defense // UPF // Feb 16th 2009 the Long Tail model [Kilkki, 2007] • Split the curve in three parts (82 artists) (6,573 artists) (~254K artists)
  56. PhD defense // UPF // Feb 16th 2009 contributions + Long Tail analysis
  57. PhD defense // UPF // Feb 16th 2009 artist indegree vs. artist popularity • Are the network hubs the most popular artists? ???
  58. PhD defense // UPF // Feb 16th 2009 artist indegree vs. artist popularity Last.fm: correlation between Kin and playcounts   r = 0.621
  59. PhD defense // UPF // Feb 16th 2009 artist indegree vs. artist popularity Audio CB similarity: no correlation   r = 0.032
  60. PhD defense // UPF // Feb 16th 2009 artist indegree vs. artist popularity Expert: correlation between Kin and playcounts   r = 0.475
  61. PhD defense // UPF // Feb 16th 2009 navigation along the Long Tail • “From Hits to Niches” # clicks to reach a Tail artist, starting in the Head  how many clicks?
  62. PhD defense // UPF // Feb 16th 2009 navigation along the Long Tail • “From Hits to Niches” Audio CB similarity example (VIDEO) 
  63. PhD defense // UPF // Feb 16th 2009 navigation along the Long Tail • “From Hits to Niches” Audio CB similarity example   Bruce Springsteen (14,433,411 plays)
  64. PhD defense // UPF // Feb 16th 2009 navigation along the Long Tail • “From Hits to Niches” Audio CB similarity example   Bruce Springsteen (14,433,411 plays)  The Rolling Stones (27,720,169 plays)
  65. PhD defense // UPF // Feb 16th 2009 navigation along the Long Tail • “From Hits to Niches” Audio CB similarity example   Bruce Springsteen (14,433,411 plays)  The Rolling Stones (27,720,169 plays)  Mike Shupp (577 plays)
  66. PhD defense // UPF // Feb 16th 2009 artist similarity vs. artist popularity • navigation in the Long Tail Similar artists, given an artist in the HEAD part:  CF CB EXP 64,74% 60,92% 54,68% 45,32% 33,26% 28,80% (0%) 6,46% 5,82% Head Mid Tail Head Mid Tail Head Mid Tail Also, it can be seen as a Markovian Stochastic  process...
  67. PhD defense // UPF // Feb 16th 2009 artist similarity vs. artist popularity • navigation in the Long Tail Markov transition matrix 
  68. PhD defense // UPF // Feb 16th 2009 artist similarity vs. artist popularity • navigation in the Long Tail Markov transition matrix 
  69. PhD defense // UPF // Feb 16th 2009 artist similarity vs. artist popularity • navigation in the Long Tail Last.fm Markov transition matrix 
  70. PhD defense // UPF // Feb 16th 2009 artist similarity vs. artist popularity • navigation in the Long Tail From Head to Tail, with P(T|H) > 0.4  Number of clicks needed   CF : 5  CB : 2  EXP: 2 HEAD #clicks? TAIL
  71. PhD defense // UPF // Feb 16th 2009 artist popularity Summary |-----------------------|---------|-----|-----------| | | Last.fm | CB | Exp (AMG) | |-----------------------|---------|-----|-----------| | Indegree / popularity| yes | no | yes | | | | | | |Similarity / popularity| yes | no | no | |-----------------------|---------|-----|-----------|
  72. PhD defense // UPF // Feb 16th 2009 summary: complex networks+popularity |-----------------------|---------|-----|-----------| | | Last.fm | CB | Exp (AMG) | |-----------------------|---------|-----|-----------| | Small World | yes | yes | yes | | | | | | | Scale-free | yes | no | no | | | | | | | Ass. mixing | yes | no | no | |-----------------------|---------|-----|-----------| | Indegree / popularity| yes | no | yes | | | | | | |Similarity / popularity| yes | no | no | |-----------------------|---------|-----|-----------| | POPULARITY BIAS | YES | NO | FAIRLY | |-----------------------|---------|-----|-----------|
  73. PhD defense // UPF // Feb 16th 2009 contributions 1) Network-based evaluation Item Popularity + Complex networks 2) User-based evaluation 3) Systems
  74. PhD defense // UPF // Feb 16th 2009 contribution #2: User-based evaluation • How do users perceive novel, non-obvious recommendations? Survey   288 participants Method: blind music recommendation   no metadata (artist name, song title)  only 30 sec. audio excerpt
  75. PhD defense // UPF // Feb 16th 2009 music recommendation survey • 3 approaches: CF: Social-based Last.fm similar tracks  CB: Pure audio content-based similarity  HYbrid: AMG experts + audio CB to rerank songs   (Not a combination of the two previous approaches) • User profile: last.fm, top-10 artists  • Procedure Do you recognize the song?   Yes, Only Artist, Both Artist and Song title Do you like the song?   Rating: [1..5]
  76. PhD defense // UPF // Feb 16th 2009
  77. PhD defense // UPF // Feb 16th 2009 music recommendation survey: results • Overall results
  78. PhD defense // UPF // Feb 16th 2009 music recommendation survey: results • Overall results
  79. PhD defense // UPF // Feb 16th 2009 music recommendation survey: results • Familiar recommendations (Artist & Song)
  80. PhD defense // UPF // Feb 16th 2009 music recommendation survey: results • Ratings for novel recommendations
  81. PhD defense // UPF // Feb 16th 2009 music recommendation survey: results • Ratings for novel recommendations one-way ANOVA within subjects (F=29.13, p<0.05)  Tukey's test (pairwise comparison) 
  82. PhD defense // UPF // Feb 16th 2009 music recommendation survey: results • % of novel recommendations
  83. PhD defense // UPF // Feb 16th 2009 music recommendation survey: results • % of novel recommendations one-way ANOVA within subjects (F=7,57, p<0.05)  Tukey's test (pairwise comparison) 
  84. PhD defense // UPF // Feb 16th 2009 music recommendation survey: results • Novel recommendations Last.fm provides less % of novel songs, but of  higher quality
  85. PhD defense // UPF // Feb 16th 2009 contributions 1) Network-based evaluation Item Popularity + Complex networks 2) User-based evaluation 3) Systems
  86. PhD defense // UPF // Feb 16th 2009 Why? besides better understanding of music recommendation... Open questions in the State of the Art in music discovery & recommendation: Is it possible to create a music discovery engine exploiting the music content in the WWW? How to build it? How can we describe the available music content? => SearchSounds Is it possible to recommend, filter and personalize music content available on the WWW? How to describe a user profile? What can we recommend beyond similar artists? => FOAFing the Music
  87. PhD defense // UPF // Feb 16th 2009 contribution #3: two complete systems • Searchsounds Music search engine   keyword based search  “More like this” (audio CB)
  88. PhD defense // UPF // Feb 16th 2009 contribution #3: two complete systems • Searchsounds Crawl MP3 blogs  > 400K songs analyzed 
  89. PhD defense // UPF // Feb 16th 2009 contribution #3: two complete systems • Searchsounds Further work: improve song descriptions using   Auto-tagging [Lamere, 2008] [Turnbull, 2007] audio CB similarity [Sordo et al., 2007] tags from the text (music dictionary)  Feedback from the users thumbs-up/down tag audio content
  90. PhD defense // UPF // Feb 16th 2009 contribution #3: two complete systems • FOAFing the music Music recommendation   constantly gathering music related info via RSS feeds  It offers: artist recommendation new music releases (iTunes, Amazon, eMusic, Rhapsody, Yahoo! Shopping) album reviews concerts close to user's locations related mp3 blogs and podcasts
  91. PhD defense // UPF // Feb 16th 2009 contribution #3: two complete systems • FOAFing the music Integrates different user accounts (circa 2005!)  Semantic Web (FOAF, OWL/RDF) + Web 2.0  2nd prize Semantic Web Challenge (ISWC 2006) 
  92. PhD defense // UPF // Feb 16th 2009 contribution #3: two complete systems • FOAFing the music Further work:   Follow Linking Open Data best practices  Link our music recommendation ontology with Music Ontology [Raimond et al., 2007]  (Automatically) add external information from: Myspace Jamendo Garageband ...
  93. PhD defense // UPF // Feb 16th 2009 summary of contributions :: research questions • 1) How can we evaluate/compare different music recommendation approaches? • 2) How far into the Long Tail do music recommenders reach? • 3) How do users perceive novel (unknown to them), non-obvious recommendations?
  94. PhD defense // UPF // Feb 16th 2009 summary of contributions :: research questions • 1) How can we evaluate/compare different music recommendation approaches? Objective framework comparing music rec.  approaches (CF, CB, EX) using Complex Network analysis Highlights fundamental differences among the  approaches • 2) How far into the Long Tail do music recommenders reach? • 3) How do users perceive novel (unknown to them), non-obvious recommendations?
  95. PhD defense // UPF // Feb 16th 2009 summary of contributions :: research questions • 1) How can we evaluate/compare different music recommendation approaches? • 2) How far into the Long Tail do music recommenders reach? Combine 1) with the Long Tail model, and Markov  model theory Highlights differences in terms of discovery and  navigation • 3) How do users perceive novel (unknown to them), non-obvious recommendations?
  96. PhD defense // UPF // Feb 16th 2009 summary of contributions :: research questions • 1) How can we evaluate/compare different music recommendation approaches? • 2) How far into the Long Tail do music recommenders reach? • 3) How do users perceive novel (unknown to them), non-obvious recommendations? Survey with 288 participants  Still room to improve novelty (3/5 or less...)   To appreciate novelty users need to understand the recommendations
  97. PhD defense // UPF // Feb 16th 2009 summary of contributions :: research questions • 1) How can we evaluate/compare different music recommendation approaches? • 2) How far into the Long Tail do music recommenders reach? • 3) How do users perceive novel (unknown to them), non-obvious recommendations? => Systems that perform best (CF) do not exploit the  Long Tail, and Systems that can ease Long Tail navigation (CB) do  not perform good enough Combine (hybrid) different approaches! 
  98. PhD defense // UPF // Feb 16th 2009 Systems that perform  best (CF) do not exploit the Long Tail, and Systems that can ease  Long Tail navigation (CB) do not perform good enough Combine different  approaches!
  99. PhD defense // UPF // Feb 16th 2009 summary of contributions :: systems • Furthermore... 2 web systems that improved existing State of the  Art work in music discovery and recommendation  Searchsounds: music search engine exploiting music related content in the WWW  FOAFing the Music: music recommender based on a FOAF user profile, also offering a number of extra features to complement the recommendations
  100. PhD defense // UPF // Feb 16th 2009 further work :: limitations • 1) How can we evaluate/compare different recommendations approaches? Dynamic networks  [Leskovec, 2008]  track item similarity over time  track user's taste over time  trend and hype detection
  101. PhD defense // UPF // Feb 16th 2009 further work :: limitations • 2) How far into the Long Tail do recommendation algorithms reach? Intercollections  how to detect bad quality music in the tail? 
  102. PhD defense // UPF // Feb 16th 2009 further work :: limitations • 3) How do users perceive novel, non-obvious recommendations?  User understanding [Jennings, 2007]  savant, enthusiast, casual, indifferent Transparent, steerable recommendations  [Lamere & Maillet, 2008]  Why? as important as What?
  103. PhD defense // UPF // Feb 16th 2009 summary: articles • #1) Network-based evaluation for RS  O. Celma and P. Cano. “From hits to niches? or how popular artists can bias music recommendation and discovery”. ACM KDD, 2008.  J. Park, O. Celma, M. Koppenberger, P. Cano, and J. M. Buldu. “The social network of contemporary popular musicians”. Journal of Bifurcation and Chaos (IJBC), 17:2281–2288, 2007.  M. Zanin, P. Cano, J. M. Buldu, and O. Celma. “Complex networks in recommendation systems”. WSEAS, 2008  P. Cano, O. Celma, M. Koppenberger, and J. M. Buldu “Topology of music recommendation networks”. Journal Chaos (16), 2006. • #2) User-based evaluation for RS  O. Celma and P. Herrera. “A new approach to evaluating novel recommendations”. ACM RecSys, 2008.
  104. PhD defense // UPF // Feb 16th 2009 summary: articles • #3) Prototypes FOAFing the Music   O. Celma and X. Serra. “FOAFing the music: Bridging the semantic gap in music recommendation”. Journal of Web Semantics, 6(4):250–256, 2008.  O. Celma. “FOAFing the music”. 2nd Prize Semantic Web Challenge ISWC, 2006.  O. Celma, M. Ramirez, and P. Herrera. “FOAFing the music: A music recommendation system based on rss feeds and user preferences”. ISMIR, 2005.  O. Celma, M. Ramirez, and P. Herrera. “Getting music recommendations and filtering newsfeeds from foaf descriptions”. Scripting for the Semantic Web, ESWC, 2005.
  105. PhD defense // UPF // Feb 16th 2009 summary: articles • #3) Prototypes Searchsounds   O. Celma, P. Cano, and P. Herrera. “Search sounds: An audio crawler focused on weblogs”. ISMIR, 2006.  V. Sandvold, T. Aussenac, O. Celma, and P. Herrera. “Good vibrations: Music discovery through personal musical concepts”. ISMIR, 2006.  M. Sordo, C. Laurier, and O. Celma. “Annotating music collections: how content-based similarity helps to propagate labels”. ISMIR, 2007.
  106. PhD defense // UPF // Feb 16th 2009 summary: articles • Misc. (mainly MM semantics)  R. Garcia C. Tsinaraki, O. Celma, and S. Christodoulakis. “Multimedia Content Description using Semantic Web Languages” book, Chapter 2. Springer–Verlag, 2008.  O. Celma and Y. Raimond. “Zempod: A semantic web approach to podcasting”. Journal of Web Semantics, 6(2):162–169, 2008.  S. Boll, T. Burger, O. Celma, C. Halaschek-Wiener, E. Mannens. “Multimedia vocabularies on the Semantic Web”. W3C Technical report, 2007.  O. Celma, P. Herrera, and X. Serra. “Bridging the music semantic gap”. SAMT, 2006.  R. Garcia and O. Celma. “Semantic integration and retrieval of multimedia metadata”. ESWC, 2005
  107. PhD defense // UPF // Feb 16th 2009 summary: articles  R. Troncy, O. Celma, S. Little, R. Garcia and C. Tsinaraki. “MPEG-7 based multimedia ontologies: Interoperability support or interoperability issue?” MARESO, 2007.  M. Sordo, O. Celma, M. Blech, and E. Guaus. “The quest for musical genres: Do the experts and the wisdom of crowds agree?”. ISMIR, 2008. • Music Recommendation Tutorials -- with Paul Lamere  ACM MM, 2008 (Vancouver, Canada)  ISMIR, 2007 (Vienna, Austria)  MICAI, 2007 (Aguascalientes, Mexico)
  108. PhD defense // UPF // Feb 16th 2009 summary: dissemination • PhD Webpage http://mtg.upf.edu/~ocelma/PhD   PDF  Source code Long Tail Model in R  References Citeulike  Related links delicious
  109. PhD defense // UPF // Feb 16th 2009 acknowledgments NB: The complete list of acknowledgments can be found in the document
  110. Music Recommendation and Discovery in the Long Tail Òscar Celma Doctoral Thesis Defense (Music Technology Group ~ Universitat Pompeu Fabra)
  111. PICA-PICA UPF-Tanger, 3rd floor

×