Music Recommendation and Discovery
β’
77 likesβ’51,629 views
Tutorial on Music Recommendation by Oscar Celma (Gracenote) and Paul Lamere (The Echo Nest). The world of music is changing rapidly. We are now just a few clicks away from being able to listen to nearly any song that has ever been recorded. This easy access to a nearly endless supply of music is changing how we explore, discover, share and experience music. As the world of online music grows, music recommendation and discovery tools become an increasingly important way for music listeners to engage with music. Commercial recommenders such as Last.fm, iTunes Genius and Pandora have enjoyed commercial and critical success. But how well do these systems really work? How good are the recommendations? How far into the "long tail" do these recommenders reach? In this tutorial we look at the current state-of-the-art in music recommendation and discovery. We examine current commercial and research systems, focusing on the advantages and the disadvantages of the various recommendation strategies. We look at some of the challenges in building music recommenders and we explore some of the novel techniques that are being used to improve future music recommendation and discovery systems. Γscar Celma is the Chief Innovation Officer at Barcelona Music and Audio Technologies (BMAT). In 2008, Γscar obtained his Ph.D. in Computer Science and Digital Communication, in the Pompeu Fabra University (Barcelona, Spain). Γscar has a book published by Springer, titled "Music Recommendation and Discovery: The Long Tail, Long Fail and Long Play in the Music Digital Age" (2010). He holds 2 patents (US2003009344 and JP2003323188, 2002) from his work on the Vocaloid system, a singing voice-synthesizer bought by Yamaha in 2004. Follow on Twitter: @ocelma Paul Lamere is the Director of Developer Platform for The Echo Nest, a music intelligence company located in Boston. Paul is interested in using technology to help people explore for new and interesting music. He is active in both the music information retrieval and the recommender systems research communities. Paul authors a popular blog on music technology at MusicMachinery.com. Follow on Twitter: @plamere
Recommended
Recommended
More Related Content
Viewers also liked
Viewers also liked (20)
More from Paul Lamere
More from Paul Lamere (10)
Recently uploaded
Recently uploaded (20)
Music Recommendation and Discovery
- 1. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Music Recommendation and Discovery Remastered Tutorial @recsys, 2011
- 2. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma @plamere @ocelma
- 3. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma INTRO CHORUS VERSE BRIDGE OUTRO
- 4. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma INTRO CHORUS VERSE BRIDGE OUTRO
- 5. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Music Recommendation is important How many songs fit in my pocket? 10 Songs 1,000 Songs 10,000,000 Songs 1979 2001 2011
- 6. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma What's so special about music? β Huge item space β Very low cost per item β Many item types β Low consumption time β Very high per-item reuse β Highly passionate users β Highly contextual usage β Consumed in sequences β Large personal collections β Doesn't require our full attention β Highly Social
- 7. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Music recommendation is broken ... If you like Britney Spears you might like... ...Report on Pre-War Intelligence Let's look at some of the issues ....
- 8. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma What makes a good music recommendation? β Relevance β Novelty / Serendipity β Transparency / Trust β Reach β Context
- 9. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma What makes a good music recommendation? β Relevance β Novelty / Serendipity β Transparency / Trust β Reach β Context
- 10. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Relevance
- 11. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Relevance β cold start new or unpopular items If you like Gregorian Chants you might like Green Day
- 12. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Cold Start β New User - Enrollment
- 13. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma New User β Implicit taste data The Audioscrobbler
- 14. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Relevance β Metadata Mismatches
- 15. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Relevance β Metadata Mismatches Why?
- 16. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Relevance - The grey sheep problem
- 17. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Relevance β Cultural Mismatches
- 18. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma What makes a good music recommendation? β Relevance β Novelty / Serendipity β Transparency / Trust β Reach β Context
- 19. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Novelty and Serendipity
- 20. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Popularity Bias - The Harry Potter Effect
- 21. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma ...also known as the Coldplay effect
- 22. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Novelty / Serendipity β the enemy High stakes competitions focused on relevance can reduce novelty and serendipity
- 23. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma What makes a good music recommendation? β Relevance β Novelty / Serendipity β Transparency / Trust β Reach β Context
- 24. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma The Opacity Problem βIf you like NiN you might like Johnny Cashβ Why???
- 25. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Is this a good recommendation? If you like Norah Jones ... You might like Ravi Shankar
- 26. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Is this a good recommendation? If you like Norah Jones ... You might like her father, Ravi Shankar
- 27. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Brutal Death Metal Quiz ??????? Photo cc by Mithrandir3
- 28. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Brutal Death Metal Quiz
- 29. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Hacking the recommender
- 30. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma What makes a good music recommendation? β Relevance β Novelty / Serendipity β Transparency / Trust β Reach β Context
- 31. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Help! Iβm stuck in the head The limited reach of music recommendation 48% of recommendations Popularity 0% of recommendations 52% of recommendations 83 Artists 6,659 Artists 239,798 Artists Study by Dr. Oscar Celma - MTG UPF Sales Rank
- 32. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Music Discovery Challenge Personal discovery a challenge too Listener Study Listeners 5,000 Average Songs 3,500 Per User Percent of songs never 65% listened to
- 33. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma What makes a good music recommendation? β Relevance β Novelty / Serendipity β Transparency / Trust β Reach β Context
- 34. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Music Recommendation is not just shopping β It is not just for shopping, but... β Discovery β Exploration β Play β Organization β Playlisting β Recommendation for groups β Devices β Doesn't have to look like a spreadsheet!
- 35. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Context: Tools for exploration Ishkur's Guide to Electronic Dance Music http://techno.org/electronic-music-guide/
- 36. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma INTRO CHORUS VERSE BRIDGE OUTRO
- 37. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma
- 38. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma
- 39. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Ingestion process β Input data source β Own data, Customer, Labels, UGC, ... β Protocol β Ingestion format β TSV, XML, DDEX, XLS!, β¦ β Method β FTP, API, ... β Frequency β Offline processing: Daily / weekly? β Data freshness! β Documentation
- 40. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Ingestion process β Post-processing β Data cleaning: Duplicates, normalization β Allow customer to use its own Ids! β Add links to external sources β Rosetta Stone
- 41. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Ingestion process β Considerations β Allow customer to use its own IDs when using the rec. system. β How long does it take to process the whole collection? β Incremental updates
- 42. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma
- 43. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma
- 44. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma music recommendation approaches β Expert-based β Collaborative filtering β Social-based β Content-based β Hybrid (combination)
- 45. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma music recommendation approaches β Expert-based βX similar to (or influenced by) Yβ Editorial metadata (Genre, Decades, Location, β¦) Music Genome β Collaborative filtering β Social-based β Content-based β Hybrid (combination)
- 46. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma music recommendation approaches β Expert-based β Collaborative filtering βpeople who listen to X also listen to Yβ β Social-based β Content-based β Hybrid (combination)
- 47. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma music recommendation approaches β Expert-based β Collaborative filtering βpeople who listen to X also listen to Yβ β Social-based β Content-based β Hybrid (combination)
- 48. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma music recommendation approaches β Expert-based β Collaborative filtering βpeople who listen toRawalso listen to Yβ X plays: β Social-based β Content-based β Hybrid (combination)
- 49. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma music recommendation approaches β Expert-based β Collaborative filtering βpeople who listen toRawalso listen to Yβ X plays: β Social-based Normalize to [5..1] β Content-based β Hybrid (combination)
- 50. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma music recommendation approaches β Expert-based β Collaborative filtering βpeople who listen toRawalso listen to Yβ X plays: β Social-based Normalize to [5..1] Probability distribution: β Content-based 0.32 , 0.24 , 0.18 , 0.1 , 0.07 , 0.04, 0.03 , 0.02 β Hybrid (combination)
- 51. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma music recommendation approaches β Expert-based β Collaborative filtering βpeople who listen toRawalso listen to Yβ X plays: β Social-based Normalize to [5..1] Probability distribution: β Content-based 0.32 , 0.24 , 0.18 , 0.1 , 0.07 , 0.04, 0.03 , 0.02 β Hybrid (combination) Binary: 100100000100000011000001
- 52. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma music recommendation approaches β Expert-based β Collaborative filtering Matrix Factorization. E.g: SVD, NMF, ... β Social-based β Content-based β Hybrid (combination)
- 53. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma music recommendation approaches β Expert-based β Collaborative filtering β Social-based β WebMIR [Schedl, 2008] Content Reviews Lyrics Blogs Social Tags Bios Playlists β Content-based β Hybrid (combination)
- 54. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma music recommendation approaches β Expert-based β Collaborative filtering β Social-based β Content-based βX and Y sound similarβ β Hybrid
- 55. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma music recommendation approaches β Expert-based β Collaborative filtering β Social-based β Content-based Audio features β Bag-of-frames (MFCC) [Aucouturier, 2004], Rhythm [Gouyon, 2005], Harmony [Gomez, 2006], ... Similarity β KL-divergence: GMM [Aucouturier, 2002] β EMD [Logan, 2001] β Euclidean: PCA [Cano, 2005] β Cosine: mean/var (feature vectors) β Ad-hoc
- 56. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma music recommendation approaches β Expert-based β Collaborative filtering β Social-based β Content-based Audio features β Bag-of-frames (MFCC) [Aucouturier, 2004], Rhythm [Gouyon, 2005], Harmony [Gomez, 2006], ... Similarity β KL-divergence: GMM [Aucouturier, 2002] β EMD [Logan, 2001] β Euclidean: PCA [Cano, 2005] http://xkcd.com/26/ β Cosine: mean/var (feature vectors) β Ad-hoc
- 57. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma music recommendation approaches β Expert-based β Collaborative filtering β Social-based β Content-based β Hybrid Weighted (linear combination) β E.g CF * 0.2 + CT * 0.4 + CB * 0.4 Cascade β E.g 1st apply CF, then reorder by CT or CB Switching ...
- 58. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma
- 59. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma
- 60. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Search β Metadata search Bruce* β Using filters: βPopular Irish bands from the 80sβ popularity:[8.0 TO 10.0] AND iso_country:IE AND decade:1980 β Audio search (and similarity) β Query by example
- 61. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Similarity
- 62. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Similarity ? ? ? Using Last.fm-360K dataset
- 63. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Similarity Using Last.fm-360K dataset
- 64. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Similarity' (include feedback) Using Last.fm-360K dataset
- 65. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Beyond similarity
- 66. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Recommendation β βIf Paul likes Radiohead he might also like Xβ vs. β βIf Oscar likes Radiohead he might also like Yβ
- 67. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Recommendation β βIf Paul likes Radiohead he might also like Xβ vs. β βIf Oscar likes Radiohead he might also like Yβ SIMILARITY != RECOMMENDATION
- 68. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Recommendation β To whom are we recommending? Phoenix-2 (UK, 2006)
- 69. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma lamere @ last.fm
- 70. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma mini-lamere's @ last.fm β Clustering (k-means) lamere top-50 artists
- 71. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma mini-lamere's @ last.fm β Clustering (k-means) lamere top-50 artists
- 72. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma @lamere - Radiohead
- 73. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma @lamere - Radiohead Vs. Radiohead similar artists...
- 74. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma @lamere - Radiohead Vs. Radiohead similar artists...
- 75. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma @lamere - Radiohead Vs. Radiohead similar artists...
- 76. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma @lamere - Radiohead Vs. Radiohead similar artists...
- 77. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Personalization (Itemization?) β ...but also which Radiohead era?
- 78. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Analytics β Big data processing β capture, storage, search, share, analysis and visualization β (local) Trend detection β Tastemakers β ...
- 79. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma INTRO CHORUS VERSE BRIDGE OUTRO
- 80. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Real-world Music Recommendation
- 81. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Real-world Music Recommendation
- 82. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Contextual Web Crawl
- 83. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Audio Processing
- 84. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Hybrid Recommendation
- 85. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma β 100 million registered users β 37 million active monthly users β More than 900,000 songs in catalog β More than 90,000 artists in catalog β More than 11 billion thumbs β More than 1.9 billion stations β 95% of the collection was played in July 2011
- 86. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Curation and Analysis
- 87. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Weighting vectors
- 88. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma For unknown artists
- 89. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma For popular artists
- 90. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma
- 91. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma
- 92. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Country: UK Record Labels: Acid Jazz, Sony BMG, Columbia Genres: Funk, Disco, Acid Jazz, Jazz Fusion, Pop-rock Years active: 1992 - present Associated acts: Brand New Heavies, Guru, Julian Perretta
- 93. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Country: UK Record Labels: Acid Jazz, Sony BMG, Columbia Genres: Funk, Disco, Acid Jazz, Jazz Fusion, Pop-rock Years active: 1992 - present Associated acts: Brand New Heavies, Guru, Julian Perretta Mood: upbeat, energetic Rhythm: 120bpm, no rubato, high percusiveness Key: Dm Tags: acid jazz funk dance Sounds like: Sereia (Tiefschwarz Radio Edit) by Mundo Azul
- 94. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma βI want some upbeat songs from unknown US bands, similar to Radioheadβ http://ella.bmat.ws/collections/bmat/artists/radiohead/similar/tracks ?filter=mood:happy +speed:fast +iso_country:US +popularity:[0.0+TO+4.0]
- 95. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma INTRO CHORUS VERSE BRIDGE OUTRO
- 96. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Evaluation "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." β "Introduction to Information Retrieval" (Manning, Raghavan, and Schutze, 2008)
- 97. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma RMSE
- 98. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma RMSE?
- 99. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma NO RMSE
- 100. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma NO RMSE (in music)
- 101. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Evaluation β Limitations of current metrics (RMSE, P/R, ROC, Spearman Rho, Kendall Tau, etc.) β skewness β performed on test data that users chose to rate β do not take into account β usefulness β novelty / serendipity β topology of the (item or user) similarity graph β ...
- 102. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Evaluation If no RMSE then...? β Predictive Accuracy vs. Perceived Quality β Does the recommendation help the user? (user satisfaction) β Familiarity vs. Novelty β Does the recommendation help the system? β $$$ β Catalog exposure
- 103. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma NEXT SONG?
- 104. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma NEXT SONG? ? Mean Reciprocal Rank + User feedback
- 105. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma ? ?
- 106. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma
- 107. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Novelty & Relevance
- 108. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Novelty & Relevance
- 109. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Novelty & Relevance
- 110. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Novelty & Relevance
- 111. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Novelty & Relevance WTF?
- 112. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Novelty & Relevance Emitt Rhodes
- 113. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Novelty & Relevance WHY as important as WHAT
- 114. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Novelty & Relevance WHY as important as WHAT
- 115. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Novelty & Relevance WHY as important as WHAT
- 116. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Novelty & Relevance WHY as important as WHAT
- 117. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma WTF
- 118. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Other evaluation techniques rd How can I evaluate a 3 party recommender: objective measures: coverage, reach subject measures: Focus on precision Measure irrelevant results: The WTF test
- 119. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma The WTF Test Why the Freakomendation?
- 120. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Evaluation β Research Datasets β Million Song Dataset (CB, Social, Lyrics, Tags and more) http://labrosa.ee.columbia.edu/millionsong/ β Last.fm (CF) http://ocelma.net/MusicRecommendationDataset/ β Last.fm 360K users <user, artist, total plays> β Last.fm 1K users <user, timestamp, artist, song>
- 121. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma
- 122. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma
- 123. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Monitoring β Do not monitor (or test) only the Algorithm, but the WHOLE recommender system: KPIs β Catalog β % matches against full catalog? β Ingestion time? β Availability? β Data & Algorithms β Time computing (e.g. Matrix factorization)? β Matrix size (e.g. ~10M x ~1M) in memory? β 10M vectors with 300 floats per vector β ~11Gb β Time computing vector similarity O(n)?
- 124. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Monitoring USAGE β Search assert_equal(ID(search('The The')), ID('The The')) β Similarity assert(similarity(U2, REM) > 0.8) assert(similarity(AC/DC, Rebecca Black) < 0.3) β Recommendation 0) create_profile(@ocelma) 1) assert(similarity(@ocelma, U2) >= 0.8) 2) dislike(@ocelma, track(U2,Lemon)) 3) assert(similarity(@ocelma, album(U2,Zooropa)) < 0.8)
- 125. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Monitoring β (web) API β Measure query response β Jmeter, Apache Benchmark β Process real logs β Fake (repeated) queries β fast because using cache?
- 126. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Monitoring β (web) API β Measure query response β Jmeter, Apache Benchmark β Process real logs β Fake (repeated) queries β fast because using cache?
- 127. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma INTRO CHORUS VERSE BRIDGE OUTRO
- 128. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Conclusions β Music Recsys is multidisciplinary β search and filtering, musicology, data mining, machine learning, personalization, social networks, text processing, complex networks, user interaction, information visualization, and signal processing (among others!) β Music Recsys is important β These technologies will be integral in helping the next generation of music listeners find that next favorite song β Strong industry impact β Music Recsys is special
- 129. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Further research β How well do music recommenders work? β lack of standardized data sets and objective evaluation methods β How to recognize and incorporate context into recommendations? β listenerβs context (exercising, exploring, working, driving, relaxing, and so on)
- 130. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Further research β How to make recommendations for all music? β consider all music including new, unknown, and unpopular content. β What effect will automatic music recommenders have on the collective music taste?
- 131. recsys 2011 | music recommendation and discovery tutorial | paul lamere & oscar celma Music Recommendation and Discovery Remastered Tutorial @recsys, 2011