Information organization in recommendation Rachel Bloch Shapiro College of Information Studies, University of Maryland

Echo Nest – Hybrid Approach Introduction Literature review Combining automatic content analysis with cultural Conclusion information via web crawling, the Echo Nest focuses on song level analysis, and taps into pre- Research and initiative from the academic Music recommendation presents an Researchers have proposed exciting new community and commercial ventures lead to interesting case study for new approaches ways to improve music recommendation. existing information like reviews. Songs are analyzed more quickly, and at lower cost than projects like the Million Song Dataset, and to information organization. Traditional Minimal Metadata: Bogdanov and advance recommendation. Linked data, still in its searches are not always fruitful, and Herrera propose that automated content‐ Pandora, allowing the Echo Nest to amass a huge catalog of songs. infancy, offers promise. Well-designed enabling discovery requires creativity. based recommendations, simply filtered recommendation systems will be needed in the Without the limitations of shelf space, by genre, outperform both purely content- future to ensure users are connected with relevant online retailers offer many more titles, based and programs materials. creating a business of “selling more of and require much less data storage. Discussion less” by dipping into the “,” a Crowdsourcing ‘Your head is my tail’: term popularized by Chris Anderson. Lee and Lee suggest a creative approach Music Information’s Future: Linked Data Despite millions of songs available, to the problem of popularity bias by Passant’s dbRec program generates novel approximately 90% of sales came from personalizing collaborative filtering recommendations using linked data resources such Literature cited only 2% of albums in 2011. algorithms: Identify “experts” and as dbPedia. The seevl YouTube plug in builds upon 1. Barrington, L., Oda, R. & Lanckreit, G. (2009). Smarter Than Genius? Human Evaluation of Music “novices” based on listening habits and this research and brings linked data to a huge audience. There are numerous other linked data Recommender Systems. Retrieved October 24, 2012 favorite songs and use “experts” in from http://cosmal.ucsd.edu/cal/pubs/Barrington‐ particular songs/artist to create projects, such as BBC Music, which is now Genius‐ISMIR09.pdf recommendations for novices in that area. available in RDF, and high level semantic 2. Bertin‐Mahieux, T., Ellis, D.P.W., Whitman, B. & Improving Folksonomies: We can’t rely prediction (e.g. mood) programs. Lamere, P. (2011) The Million Song Dataset. Retrieved upon the “wisdom of crowds” without a October 24, 2012 from http://ismir2011.ismir.net/ papers/OS6-1.pdf crowd. Eck and others solve the cold start 3. Bogdanov, D. & Herrera, P. (2011). How much metadata by assigning tags to untagged songs using do we need in music recommendation? A subjective automated listening and prediction. evaluation using preference sets. Retrieved October 24, Syndicated Content: Celma and others 2012 from http://ismir2011.ismir.net/papers/PS1-10.pdf mined the information syndicated in RSS 4. Celma, O.(2010) Music recommendation and discovery: The long tail, long fail and long play in the digital feeds from music sites, taking advantage Figure 1.The Long Tail distribution music space. Berlin: Springer. of rich knowledge already on the web. 5. Eck, D., Lamere, P., Mahieux, T. & Green, S. (2007). Automatic Generation of Social Tags for Music Recommendation. Retrieved October 24, 2012 from Key terms: http://www.columbia.edu/~tb2332/Papers/nips07.pdf -collaborative filtering is used by many 6. Lamere, P. & Celma, O. (2011) Music recommendation sites, “people who liked x also liked y” Case studies and discovery remastered. Retrieved October 24, 2012 -content based systems use information iTunes Genius- Collaborative Filtering from http://recsys.acm.org/2011/tutorials.shtml. about the music itself (tempo, pitch, etc) Genius’s algorithm appears to rely on 7. Lee, K. & Lee, K. (2011). My head is your tail: Applying link analysis on long‐tailed music listen-ing behavior to inform recommendations collaborative filtering and does not do any Figure 2. Part of the Linked Open Data for music recommendation. Proceedings of the Fifth -popularity bias: extremely popular items content analysis. Barrington, et al showed cloud showing music sites. ACM Conference on Recommender Systems (pp 213‐ are over-recommended (the “Harry Potter” that if the ID3 tag was removed, Genius 220). New York: ACM. effect) does not make recommendations. Music Information to Scale: the Million Song 8. Passant, A. (2010). dbrec: Music recommendations using DBpedia? Retrieved October 24, 2012 from http:// -cold start: new or unknown artists lack Pandora – Content Based Dataset fans and will not be recommended by a Pandora’s “Music Genome Project” semanticweb.org/ Previously, non-commercial research was limited collaborative filtering algorithm consists of information gathered and by the use of researchers own music collection, or curated by hand, specialists catalog each music in the public domain. The Million Song song according to ~ 400 parameters. Dataset consists of sophisticated data and metadata For further information While 95% of the songs in the catalog are for one popular million songs from the Echo Nest Please email [email protected] for the full played monthly, repetition is a problem, Catalog and is freely available. paper. due to it’s relatively small catalog.