Using Song Social Tags and Topic Models to Describe and Compare Playlists

Abstract

Playlists are a natural delivery method for music recommendation and discovery systems. Recommender systems o ering playlists must strive to make them relevant and enjoyable. In this paper we survey many current means of generating and evaluating playlists. We present a means of comparing playlists in a reduced dimensional space through the use of aggregated tag clouds and topic models. To evaluate the tness of this measure, we perform prototypical retrieval tasks on playlists taken from radio station logs gathered from Radio Paradise and Yes.com, using tags from Last.fm with the result showing better than random performance when using the query playlist's station as ground truth, while failing to do so when using time of day as ground truth. We then discuss possible applications for this measurement technique as well as ways it might be improved.

Keywords

• LDA
• Topic Models
• playlists
• music
• similarity
• information retrieval
• metric space
• social tags

Categories and Subject Descriptors

H.5.5 [Sound and Music Computing]: Signal analysis, synthesis, and processing; H.5.1 [Multimedia Information Systems]: Evaluation/methodology

1. INTRODUCTION

Inherent to the design of any recommender or retrieval system is a means of display or delivery of selected content. For a system that recommends music this means playback of an audio le. Listening to or playing a piece of music take the length time of that piece of music. Given this link between music and time, when considering what information is relevant for a recommendation it is vital to consider the context of time; that is, what music has been played before WOMRAD 2010 Workshop on Music Recommendation and Discovery, colocated with ACM RecSys 2010 (Barcelona, SPAIN) Copyright c . This is an open-access article distributed under the terms of the Creative Commons Attribution License 3.0 Unported, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. or will be played after the current recommended song. Yet little is understood about how playback order a ects the success or failure of a recommendation of a piece of music. Whether a system makes user-based, object-based or hybrid recommendations, a better awareness and use of playback order will yield an improved music recommender system.

In order to take advantage of the e ect of playback order, it is necessary to have some means of comparing playlists with one another. While ratings-based generic recommender strategies could be employed, such techniques could only be used in systems which allow for the rating of playlists directly (as opposed to the much more common rating of member songs).Alternatively, n distance measure between playlists can be used to facilitate the prediction and generation of well-ordered lists of song sequences for recommendation. This has the advantage being applicable to the vast majority of existing playlist generation systems, many of which do not to collect playlist level ratings from their users. Further, a measure of playlist distance has a number of other applications in music recommender and discovery systems including label propagation, predictive personalization and context tuning to name a few.

In this paper we propose an objective distance measure between playlists. To better understand why such a measure is needed, Section 2 provides background information in existing playlist generation and evaluation techniques. While any su ciently expressive and low-dimensional feature is compatible with our playlist measure, we use a novel social tag-based feature in this paper. This song-level feature is detailed in Section 3. This is followed by an explanation of our distance measurement itself in Section 4. Putting this into practice, we detail some proof of concept evaluation in Section 5. We discuss the results of this evaluation and possible extensions in Section 6. 2.

PLAYLIST AS DELIVERY MECHANISM

In this section we survey the use of playlists in the delivery of content in existing recommendation and retrieval systems. This is followed by a review of current evaluation methods for generated playlists. These two survey points will show both the widespread use of playlist generation in music recommendation and discovery systems and the need for more quality evaluation of these systems.

While this brief survey is focused on automatic playlist generation, there is a wealth of both academic and lay work discussing various aspects manual human-driven playlist construction that may be of interest to the reader. Work in this area tends to deal with radio (e.g. [ 1 ]) or club and dance disc jockeys (e.g. [ 13 ]), being the two principal areas where the explicit construction of ordered lists of songs are tied to the eld. It is with these areas of manual playlist construction in mind that we will examine past e orts in both automatic playlist construction and evaluation techniques. 2.1

Usage in the Wild

There have been many music recommendation and retrieval systems that employ some kind of automatic playlist construction within their system. Frequently this is done as a means of content delivery or, less often, as a way of facilitating human evaluation of an underlying process such as content-based music similarity or recommendation. What follows is a brief survey of existing methods of playlist generation both with and without human intervention.

A web based system for personalized radio is detailed in [ 20 ]. In this early system users create and publish playlists facilitated through a process analogous to collaborative ltering. This results in quasi-automatic playlist creation, with any sequence ordering depending entirely on the user. Another variation of the social interaction intermediary is shown in [ 27 ], which presents the Jukola system. This system creates playlists via democratic vote on every song using mobile devices of listeners in the same physical space. Furthering the ideas of collaborative human generation, [ 25 ] shows a system called Social Playlist. This system is based on the idea of social interaction through playlist sharing, integrating mobile devices and communal playback.

A fully automatic rule-based system is described in [ 2 ]. This system uses existing metadata such as artist name, song title, duration and beats per minute. The system is designed from the ground up to be scalable and is shown to work given a database of 200000 tracks. An approach that is derived from recommender systems is seen in [ 4 ]. Here the authors use the ratings and personalization information to derive radio for a group. An attempt to optimize a playlist based on known user preference as encoded in song selection patterns is shown in [ 30 ]. This e ort uses Gaussian process regression on user preference to infer playlists. The system uses existing a priori metadata as the features for selection. A means of using webmining derived artist similarity with content-based song similarity is used to automatically generate playlists in [ 22 ]. This system combined these two spaces in such a way as to minimize the use of signal analysis. A byproduct of this optimization is improved playlist generation as is shown in a small evaluation with human listeners.

The Poolcasting system is detailed in [ 5, 6 ]. Poolcasting uses dynamic weighting of user preferences within a group of users who are all listening to a common stream with the goal of minimizing displeasure across the entire group. This results in a system that is very similar to popular commercial radio in terms of its output. A method for created playlists using an artist social graph, weighted with acoustic similarity is shown in [ 17 ]. This method takes a start and end song and constructs a playlist using maximum ow analysis on the weighted graph. Another technique for playlist construction based on the selection of paths between the start and end songs is shown in [ 18 ]. In this system content-based similarity is used to project a set of songs onto a 2-D map, then a path is found from the start song to the end song with the goal of minimizing the step size between each member song. A recent approach uses co-occurrence in n-grams extracted from the internet radio station Radio Paradise1 to deform a content-based similarity space [ 26 ]. This deformed space is then used in a manner that is similar to [ 18 ] to generate paths from one song to another, minimizing step distance throughout the path.

Also of note is [ 31 ], which in contrast to most of the previous systems, uses nearest neighbor co-occurrence in radio playlist logs to determine song similarity. While the evaluation was preliminary this method shows promise. 2.2

Evaluation Methods

The most prevalent method of evaluation used in playlist generation systems is direct human evaluation by listening. The system detailed in [ 29 ], a rule-based automatic playlist generator that uses features derived from metadata, is similar to [ 2, 30 ]. Of note in [ 29 ] is the thorough human listener testing which shows the automatic playlist generator performing considerably better than songs ordered randomly. This evaluation, though better than most, still fails to compare the automatic playlists against human expert playlists. Additionally, to reduce test time, the evaluation uses arbitrary one minute clips from the songs rather than the entirety of the song or an intentionally chosen segment. A content-based similarity playlist generator with a novel evaluation is seen in [ 28 ]. Here the authors track the number times the user presses the skip button to move on from the currently playing song. All songs that are skipped are considered false positives and those that are completely played are treated as true positives. From this many standard information retrieval techniques can be used in the evaluation, resulting in a rich understanding of the results. Ultimately, it is still human user listening evaluation though and its biggest drawback is playback time. Assuming an average song length of ve minutes it would take an an hour and 40 minutes (per listener) to listen to 20 songs with no time for the skipped songs. This skip-based evaluation framework is further used in [ 12 ] where existing last.fm user logs (which include skip behavior) are analyzed using fuzzy set theory to determine playlist generation heuristics in the system. Additionally, many systems of playlist generation lack formal evaluation all together. 2.3

Summary

While a number of techniques have been employed to create playlists for a variety of functions, there exist limited techniques in the evaluation of generated playlists. These evaluation techniques rely heavily on time consuming human evaluation. Beyond that, there is no studied means to objectively compare one playlist with another. In Section 4 we will propose just such a means. First we will describe a novel song level feature based on tags. A tag-based feature will encode socio-cultural data that is missing from analogous content-based features, though social tags bring about some other problems.

TOPIC-MODELED TAG-CLOUDS

In order to encode playlists in a low dimensional representation we must rst represent their member songs in as a low dimensional vector. Here we use a Topic-Modeled Tag Cloud (TMTC) as a pseudo-content-based feature, in a way 1http://radioparadise.com that is functionally analogous to various pure content-based methods. Using tags and topic models in this way is novel and what follows is an explanation of the process of building this feature.

A tag is a word or phrase used to describe a document of some kind, typically on the Web. Various kinds of documents are described using tags on the Web including photos2, videos3 and music4. An aggregated collection of tags, weighted by the number of users who ascribe it to a given object, is commonly referred to as a tag cloud.

Tag clouds get their name from the most common visualization method used with them, where each tag is displayed with the font size in proportion to the weight, arranged in a way that resembles a cloud. An example of a tag cloud5 can be seen in Figure 1 As can be seen in this example, tag clouds provide a rich description of the music it describes. Tags and collections of tags in various forms provide the basis for many techniques within music informatics including recommendation, retrieval and discovery applications [ 3,23 ].

In addition to human generated tags being used, there is some research directed toward the automatic application of tags and inference of associated weights on unlabeled pieces of music [ 7, 9, 16, 21 ]. 3.2

Reducing the Dimensionality

There exist some techniques (such as [ 8 ]) to determine semantic clustering within a tag cloud; however, these systems are built to facilitate browsing and do not create a su ciently reduced dimensional representation. The previous work of [ 24 ] comes the closest to the needed dimensional reduction, also dealing with social tags for music. This work, through the use of aspect models and latent semantic analysis, brings the dimensionality down into the hundreds, while preserving meaning. This order of dimensions is still too high to compute meaningful distance across multisong playlists. A feature with dimensionally of the order 102 would su er from the curse of dimensionality [ 33 ]: because of its high dimensionality, any attempt to measure distance becomes dominated by noise. However, a technique developed for improved modelling in text information retrieval, topic models provide the reduced dimensional representation 2e.g. http://flickr.com 3e.g. http://youtube.com 4e.g. http://last.fm or http://musicbrainz.org 5This tag cloud is for the track Bohemian Crapsody by the artist Sickboy. The tags and the rendering both come from last.fm, available at http://www.last.fm/music/Sickboy/ _/Bohemian+Crapsody/+tags we require. Topic models are described in [ 10 ] as \probabilistic models for uncovering the underlying semantic structure of [a] document collection based on a hierarchical Bayesian analysis of the original text." In topic modeling, a document is transformed into a bag of words, in which all of the words of a document are collected and the frequency of the occurrence in recorded. We can use the weighted collection of tags in a tag cloud as this bag of words, with tags serving as tokenized words.

There are a few di erent ways of generating topic models; for our feature generation we will be using latent Dirichlet allocation [ 11 ], treating each tag cloud as a bag-of-words. In LDA, documents (in our case tags clouds of songs) are represented as a mixture of implied (or latent ) topics, where each topic can be described as a distribution of words (or here, tags).More formally give the hyper-parameter , and the conditional multinomial parameter , Equation 3.2 gives the joint topic distribution , a set of N topics z and a set of N tags w.

N p( ; z; wj ; ) = p( j ) Y p(znj )p(wnjzn; ) n=1 (1) In Figure 2 LDA is shown as a probabilistic graphical model. In order to create topic models using LDA, we need to specify p( j ) and p(znj ). We estimate our parameters empirically from a given corpus of tag clouds. This estimation is done using variational EM as described in [ 11 ].This allows topic distributions to be generated in an unsupervised fashion, though the number of topics in a corpus must be speci ed a priori.

Once the LDA model is generated, it is used to infer the gather tags for all songs create LDA model describing

topic distributions infer topic mixtures for all

songs create vector database

of playlists mixture of topics present in the tag cloud for a given song. This is done via variational inference which is shown in [ 11 ] to estimate the topic mixture of a document by iteratively minimizing the KL divergence from variational distribution of the latent variables and the true posterior p( ; zjw; ; ).

This process in it's entirety is shown as a block diagram in Figure 3. Once this process is completed for every song in our dataset, we will have a single vector with a dimensionality equal to the number of topics in our LDA whose entries indicate topic occupancy for that song.

PLAYLISTS AS A SEQUENCE OF TOPIC

WEIGHTS

Given the single vector per song reduction, we represent the playlists these song are in as ordered sequences of these vectors. Thus each playlist is represented as a l d-dimensional vector, where l is the number of songs in a given playlist and d is the number of topics in our LDA model. 4.1

Measuring Distance

To both manage and measure the distance between these li d dimensional vectors we use audioDB6. The use of audioDB to match vectors of this type is detailed in [ 32 ]. Brie y, distance is calculated by means of a multidimensional Euclidian measure. Here li is an arbitrary length subsequence of i vectors. In practice, i is Casey:2008selected to be less than or equal to the smallest sequence length for a 6source and binary available at http://omras2.doc.gold. ac.uk/software/audiodb/ complete playlist in a dataset. The distance between two playlists is then the minimum distance between any two length i sub-vectors drawn from each playlist. One e ect of this technique is easy handling of playlists of unequal length.

This type of distance measurement has been used with success on sequences of audio frames [ 14, 15 ]. The distance measure in use between vectors can also be changed. In particular there has been work showing that statistical features (such as topic models) may bene t from the use of Manhattan distance [ 19 ], however for our prototypical evaluation we have used simple Euclidean distance as seen in equation ?? above. 5.

EVALUATION

The goal of our evaluation is to show the tness of our distance measurement through preliminary retrieval tests: searching for playlists that start at the same time of day as our query playlist and searching for the playlists from the same station from a database of stations of the same genre. We examine the logs of a large collection of radio stations, exhaustively searching example sets. Through precision and recall we see that our measure organizes playlists in a predictable and expected way. 5.1

Dataset

In order to test these proposed techniques a collection of radio station logs were gathered. These logs come from a collection of broadcast and online stations gathered via Yes.com7. The logs cover the songs played by all indexed stations between 19-26 March 2010. For our evaluation task using this data source we looked at subsets of this complete capture, based on genre labels applied to these stations. Speci cally we examine stations of the genres rock and jazz. The complete Yes.com dataset also includes stations in the following genre categories: Christian, Country, Electronica, Hip-Hop, Latin, Metal, Pop, Punk, R&B/Soul, Smooth Jazz and World. These labels are applied by the stations themselves and the categories are curated by Yes.com. Additionally, the play logs from Radio Paradise8 from 1 January 2007 to 28 August 2008 form a second set. We then attempted to retrieve tag clouds from Last.fm9 for all songs in these logs. When tags were not found the song and its associated playlist were removed from our dataset

These logs are then parsed into playlists. For the radio logs retrieved via the Yes api, the top of every hour was used as a segmentation point as a facsimile for the boundary between distinct programs. This is done under the assumption that program are more likely than not to start and nish on the hour in US commercial broadcast. Note that this method of boundary placement will almost certainly oversegment radio programs as many radio programs are longer than one hour. However, given that our distance measure compares xed length song sequences across playlists, this over-segmentation should produce only minimal distortion in our results. The Radio Paradise logs include all the links or breaks between songs where the presenter speaks brie y. For experiments using the Radio Paradise logs these links are used as playlist boundaries. This leads to a slight di erence in the type of playlist used from Radio Paradise versus Yes. 7http://api.yes.com 8http://www.radioparadise.com/ 9http://last.fm source 2543 865 1092 2246

The playlists coming from Radio Paradise represent strings of continuously played songs, with no breaks between the songs in the playlists. The playlists from Yes are approximations of a complete radio program and can therefore contain some material inserted between songs (e.g. presenter link, commercials).

Statistics for our dataset can be see in Table 1 we then use the tags clouds for these songs to estimate LDA topic models as described in Section 310. For all our experiments we specify 10 topic models a priori. The ve most relevant tags in each of the topics in models trained on both the rock and jazz stations can be seen Table 2.

Our rst evaluation looks at the di erence between the time of day a given query playlist starts and the start time for the closest n playlists by our measure. For this evaluation we looked at the 18 month log from Radio Paradise as well as the \Rock" and \jazz" labelled stations from Yes.com, each in turn. Further we used a twelve hour clock to account for The basis for this test relies on the hypothesis that for much commercial radio content in the United States, branding of programs is based on daily repeatable of tone and content for a given time of day. It should therefore be expected that playlists with similar contours would occur at similar times of day across stations competing for similar markets of listeners.

Figure 4 shows the mean across all query playlists of the time di erence for each result position for the closest n results, where n is 200 for the Radio Paradise set and 100 for the Yes.com set. The mean time di erence across all three sets is basically at, with an average time di erence of just below 11000 or about three hours. Given the maximum di erence of 12 hours, this result is entirely the opposite of compelling, with the retrieved results showing no corespondance to time of day. Further investigation is required to determine whether this is a failure of the distance metric or simply an accurate portrail of the radio stations logs. A deeper examination of some of the Yes.com data shows some evidence of the latter case. Many of the playlist queries exactly match (distance of 0) with the entirity of the 200 returned results. Further these exact match playlists are repeated evenly throughout the day. One of these queries is shown in Figure 5. The existance of these repeating playlists throughout the day, ensures this task will not con rm our 10Our topic models are created using the open source implementation of LDA found in the gensim python package available at http://nlp.fi.muni.cz/projekty/gensim/ which in turn is based on Blei's C implementation available at http://www.cs.princeton.edu/~blei/lda-c/ 25000 sd20000 n o c e s n i, se15000 m itttrs a en10000 e w t e b a t led 5000

00 hypothesis, perhaps due to progaming with no reliance on time of day, at least in the case of Radio Paradise.

Playlist beginning at midnight on 1 January 2007 50

In this evaluation we examined the precision and recall of retrieving playlists from the same station as the query playlist. Here we looked at the \Rock" and \Jazz" labelled stations retrieved via the Yes API, each in turn. Similar to the rst task, it is expected that a given station will have its own tone or particular feel that should lead to playlists from that station being more apt to match playlist from their generating station then with other stations from the same genre. More formally, for each query we treat returned playlists as relevant, true positives when they come from the same station as the query playlist and false positives otherwise. Based on this relevance assumption, precision and recall can be calculated using the following standard equations.

P = jfrelevantplaylistsg Tfretrievedplaylistsgj jfretrievedplaylistsgj R = jfrelevantplaylistsg Tfretrievedplaylistsgj jfrelevantplaylistsgj (2) (3) The precision versus recall for a selection of stations' playlists from both the \Rock" and \Jazz" stations are show in Figure 6. When considering the precision and recall performance it Bob Marley Feist john mayer drunk love feist backing vocals john mayer acoustic corinne bailey rae bonnie raitt David Pack 2 female vocalists Anna Nalick Chicas playlist 2009 Sarah McLachlan 60s jazz - sax acid jazz reggae cool jazz t8 aupa Pete whistling Triple J Hottest 100 review fun as fuck Sade deserves another listen till you come to me piano 2010 classic rock TRB 60s ElectronicaDance 70s mysterious The Beatles best songs of 2009 the rolling stones tribute to george reminds me of winter kings of leon songs that save my life songs to travel Muse

Needtobreathe

plvaronaswow2009 The Script brilliant music van morrison follow-up jazz instrumental guitar latin jazz rnb soul female vocalists Neo-Soul Robin Thicke female vocalists norah jones dido jazz vocal jazz classic rock 80s rock 70s yacht rock 80s new wave david bowie neuentd synth pop

Flamenco

tactile smooth jazz guitar ponder cafe mocha wine is useful to compare against random chance retrieval. There are 100 stations labeled \Rock" and 48 labeled \Jazz". Under chance retrieval a precision of 0:01 would be seen for \Rock" and 0:0208 for \Jazz". 5.4

Summary

Two di erent evaluation tasks have been run using real world radio log data to examine the usefulness of our playlist match technique. The rst of these, an examination the time di erence was at across result length variance. While this implies lack of discrimination into daily patterns, it is not possible to determine from the available data whether this is an accurate re ection of the progamming within the dataset or distance measure not being su cient for the task. The second task shows the performance of retrieving hourly playlists from a selection of stations using playlists from that station as a query. Here we see a great deal of promise, especially when comparing the query results against random chance, which it outperforms considerably.

CONCLUSIONS

Having reviewed recent work in various methods of playlist generation and evaluation in Section 2, it is apparent that there is a need for better ways to objectively compare playlists to one another. We detailed a method of doing so in Section 4, though rst, to better lter content-based data through listeners' experience we presented a novel tag-based feature, TMTC, using tags summarized using LDA topic models in Section 3. This was follow by two task evaluations to examine out playlist matching technique and song feature on real world playlist data from radio logs in Section 5.

While our evaluation shows the promise of this technique on sampled data, there is much room for improvement. Principal among these is the exploration of non-Euclidean distance measures. Manhattan distance (or L1) seems to have the most direct applicability and its use could prove to be quite bene cial. Another area for future work is in the use of the measure on further data and datasets. One of the best ways to improve here would be in the use of datasets with a more exact known ground truth, in order to best apply known recommender and retrieval evaluation methods to them.

This leads to a further avenue of future work, testing the measure against direct human evaluation. While our matching technique has many uses with recommendation and discovery, if it proved to align with human evaluation it would be considerably more useful. 7.

ACKNOWLEDGMENTS

This work is supported in part by the Engineering and Physical Sciences Research Council via the Online Music Recognition And Searching II (OMRAS2) project, reference number EP/E02274X/1. Additional support provided as part of the Networked Environments for Music Analysis (NEMA) project, funded by The Andrew W. Mellon Foundation. Thanks also to Paul Lamere for some dataset acquisition assistance. 8. 1.0 0.8 iion0.6 s c e rp0.4 0.2 1.0 0.8 iion0.6 s c e rp0.4 0.2 0.00.0

Radio Paradise, Start time delta v. result order YES.com Rock stations, Start time delta v. result order YES.com Jazz stations, Start time delta v. result order 14000

References

[1] J. A. Ahlkvist and R. Faulkner . ` Will This Record Work for Us?': Managing Music Formats in Commercial Radio . Qualitative Sociology , 25 ( 2 ): 189 { 215 , June 2002 .

[2] J.-J. Aucouturier and F. Pachet . Scaling up playlist generation . In Proc. IEEE International Conference 14000 snd co12000 sseen i, m it tr10000 tsan e e w t e ltab 8000 e d 6000 on Multimedia and Expo , 2002 .

[3] J.-J. Aucouturier and E. Pampalk . Introduction-from genres to tags: A little epistemology of music information retrieval research . Journal of New Music Research , 37 ( 2 ): 87 { 92 , 2008 .

[4] P. Avesani , P. Massa , M. Nori , and A. Susi . Collaborative radio community . In AH '02: Proceedings of the Second International Conference on Adaptive Hypermedia and Adaptive Web-Based Systems , pages 462 { 465 . Springer Berlin / Heideberg, January 2002 .

[5] C. Baccigalupo . Poolcasting: an intelligent technique to customise music programmes for their audience . PhD thesis , Institut d'Investigacio en Intelligencia Arti cial, 2009 .

[6] C. Baccigalupo and E. Plaza . Sharing and combining listening experience: A social approach to web radio . In Proc. of the International Computer Music Confernce , 2007 .

[7] L. Barrington , D. Turnbull , and G. Lanckriet . Auto-tagging music content with semantic multinomials . In Proc. of Int. Conference on Music Information Retrieval , 2008 .

[8] G. Begelman , P. Keller, and F. Smadja . Automated Tag Clustering: Improving search and exploration in the tag space . In In Proc. of the Collaborative Web Tagging Workshop at WWW'06 , 2006 .

[9] T. Bertin-Mahieux , D. Eck , F. Maillet , and P. Lamere . Autotagger: a model for predicting social tags from acoustic features on large music databases . Journal of New Music Research , 37 ( 2 ): 101 { 121 , 2008 .

[10] D. Blei and J. La erty. Topic Models . Text Mining: Theory and Applications . Taylor and Francis, 2009 .

[11] D. M. Blei , A. Ng , and M. Jordan . Latent dirichlet allocation . Journal of Machine Learning Research , 3 : 993 { 1022 , 2003 .

[12] K. Bosteels , E. Pampalk, and E. E. Kerre . Evaluating and analysing dynamic playlist generation heuristics using radio logs and fuzzy set theory . In Proc. of Int. Conference on Music Information Retrieval , October 2009 .

[13] B. Brewster and F. Broughton. Last Night A DJ Saved My Life; The history of the disc jockey . Headline Book Publishing, London, United Kingdom, 2nd edition , 2006 .

[14] M. Casey , C. Rhodes , and M. Slaney . Analysis of minimum distances in high-dimensional musical spaces . Audio, Speech, and Language Processing , IEEE Transactions on, 16 ( 5 ): 1015 {1028, jul. 2008 .

[15] M. Casey and M. Slaney . The importance of sequences in music similarity . In Proc. IEEE Int. Conf. on Acoustics , Speech, Signal Processing, Toulouse, France, 2006 .

[16] D. Eck , P. Lamere , T. Bertin-Mahieux , and S. Green . Automatic generation of social tags for music recommendation . In Neural Information Processing Systems Conference (NIPS) 20 , 2007 .

[17] B. Fields , K. Jacobson , C. Rhodes , and M. Casey . Social playlists and bottleneck measurements : Exploiting musician social graphs using content-based dissimilarity and pairwise maximum ow values . In Proc. of Int. Symposium on Music Information Retrieval , September 2008 .

[18] A. Flexer , D. Schnitzer , M. Gasser , and G. Widmer . Playlist generation using start and end songs . In Proc. of Int. Symposium on Music Information Retrieval , October 2008 .

[19] K. Grauman and T. Darrell . Fast contour matching using approximate earth mover's distance . In Proc. of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition , 2004 .

[20] C. Hayes and P. Cunningham . Smart radio: Building music radio on the y . In Expert Systems 2000 , pages 2{6 . ACM Press, 2000 .

[21] M. D. Ho man , D. M. Blei , and P. R. Cook . Easy as cba: a simple probabilistic model for tagging music . In Proc. of Int. Conference on Music Information Retrieval , 2009 .

[22] P. Knees , T. Pohle , M. Schedl , and G. Widmer. Combining audio-based similarity with web-based data to accelerate automatic music playlist generation . In Proc. 8th ACM international workshop on Multimedia information retrieval , pages 147 { 154 , 2006 .

[23] P. Lamere . Social tagging and music information retrieval . Journal of New Music Research , 37 ( 2 ), June 2008 .

[24] M. Levy and M. Sandler . Learning latent semantic models for music from social tags . Journal of New Music Research , 37 ( 2 ): 137 { 150 , 2008 .

[25] K. Liu and R. A. Reimer . Social playlist: enabling touch points and enriching ongoing relationships through collaborative mobile music listening . In MobileHCI '08: Proceedings of the 10th international conference on Human computer interaction with mobile devices and services , pages 403 { 406 , New York, NY, USA, 2008 . ACM.

[26] F. Maillet , D. Eck , G. Desjardins, and P. Lamere . Steerable playlist generation by learning song similarity from radio station playlists . In Proc. of Int. Conference on Music Information Retrieval , October 2009 .

[27] K. O'Hara , M. Lipson , M. Jansen , A. Unger , H. Je ries, and P. Macer . Jukola: democratic music choice in a public space . In DIS '04: Proceedings of the 5th conference on Designing interactive systems , pages 145 { 154 , New York, NY, USA, 2004 . ACM.

[28] E. Pampalk , T. Pohle , and G. Widmer . Dynamic playlist generation based on skipping behavior . In Proc. of Int. Symposium on Music Information Retrieval , 2005 .

[29] S. Pauws and B. Eggen . Pats: Realization and user evaluation of an automatic playlist generator . In Proc. of Int. Conference on Music Information Retrieval , 2002 .

[30] J. C. Platt , C. J. Burges , S. Swenson , C. Weare , and A. Zheng . Learning a gaussian process prior for automatically generating music playlists . In Proc. Advances in Neural Information Processing Systems , volume 14 , pages 1425 { 1432 , 2002 .

[31] R. Ragno , C. Burges , and C. Herley . Inferring similarity between music objects with application to playlist generation . In Proc. 7th ACM SIGMM international workshop on Multimedia information retrieval , 2005 .

[32] C. Rhodes , T. Crawford , M. Casey , and M. d'Inverno. Investigating music collections at di erent scales with audiodb . Journal of New Music Research , to appear, 2010 .

[33] R. Weber , H. J. Schek , and S. Blott . A quantitative analysis and performance study for similarity-search methods in high-dimensional spaces . In Proc. of the Intl. Conf. on Very Large Databases , 1998 .