PATS: Realization and User Evaluation of an Automatic Playlist Generator PATS: Realization and User Evaluation of an Automatic Playlist Generator Steffen Pauws Philips Research Eindhoven Prof. Holstlaan 4 (WY21) 5656 AA Eindhoven, the Netherlands +31 40 27 45415 steffen.pauws@philips.com Berry Eggen Philips Research Eindhoven, and Technische Universiteit Eindhoven / Faculty of Industrial Design Eindhoven, the Netherlands j.h.eggen@tue.nl ABSTRACT A means to ease selecting preferred music referred to as Personalized Automatic Track Selection (PATS) has been developed. PATS generates playlists that suit a particular context- of-use, that is, the real-world environment in which the music is heard. To create playlists, it uses a dynamic clustering method in which songs are grouped based on their attribute similarity. The similarity measure selectively weighs attribute-values, as not all attribute-values are equally important in a context-of-use. An inductive learning algorithm is used to reveal the most important attribute-values for a context-of-use from preference feedback of the user. In a controlled user experiment, the quality of PATS- compiled and randomly assembled playlists for jazz music was assessed in two contexts-of-use. The quality of the randomly assembled playlists was used as base-line. The two contexts-of-use were listening to soft music and listening to lively music. Playlist quality was measured by precision (songs that suit the context-of-use), coverage (songs that suit the context-of-use but that were not already contained in previous playlists) and a rating score. Results showed that PATS playlists contained increasingly more preferred music (increasingly higher precision), covered more preferred music in the collection (higher coverage), and were ratedhigher than randomly assembled playlists. 1. INTRODUCTION So far, music player functionality that has been designed for accessing and exploiting large personal music collections aims at providing fast and accurate ways to retrieve relevant music. This type of access generally requires well-defined targets. Music listeners need to instantaneously associate artists and song titles (or even CD and track numbers) with music. This is not an easy task to do, since titles and artists are not necessarily learnt together with the music [8]. In our view, selecting music from a large personal music collection is better described as a search for poorly defined targets. These targets are poorly defined since it is reasonable to assume that music listeners have no a-priori master list of preferred songs for every listening intention, lack precise knowledge about the music, and cannot easily express their music preference on-the-fly. Rather, choice for music requires listening to brief musical passages to recognize the music before being able to express a preference for it. If we take music programming on current music (jukebox) players as an example, it allows playing a personally created temporal sequence of songs in one go, once the playlist or program has been created. The creation of a playlist, however, can be a time- consuming choice task. It is hard to arrive at an optimal playlist as music has personal appeal to the listener and is judged on many subjective criteria. Also, optimality requires a complete and thorough examination of all available music in a collection, which is impractical to do so. Lastly, music programming consists of multiple serial music choices that influence each other choice criteria pertain to individual songs as well as already selected choices. A means to ease and speed up this music selection process could be of much help to the music listener. PATS (Personalized Automatic Track Selection) is a feature for music players that automatically creates playlists for a particular listening occasion (or context-of-use) with minimal user intervention [7]. This paper presents the realization of PATS and the results of a controlled user experiment to assess its performance. PATS has been realized by a decentralized and dynamic cluster algorithm that continually groups songs using an attribute-value-based similarity measure. A song refers to a recorded performance of an artist as can be found as a track on a CD. The clustering on similarity adheres to the listeners wish of coherent music in a playlist. Since it is likely that this coherence is based on particular attribute values of the songs, some attribute values contribute more than others in the computation of the similarity by the use of weights. At the same time, the clustering allows groups of songs to dissolve to form new groups. This concept adheres to the listeners wish of varied music within a playlist and over time. Clusters are presented to the music listener as playlists from which the listener can remove songs that do not meet the expectations of what a playlist should contain. An inductive learning algorithm based on decision trees is then employed that tries to reveal the attribute values that might explain the removal of songs. Weights of attribute values are adjusted accordingly, and the clustering continues with these new weights aiming at providing better future playlists. 2. PATS: EASY WAY TO SELECT MUSIC Some widely used terms such as context-of-use and music preference need further clarification. Also, we tell what we mean with minimal user intervention and explain the requirements for PATS. 2.1 Context-of-use We define context-of-use as the real-world environment in which the music is heard, being it a party, romantic evening or the traveling by car or train. The use of this concept is thought to be a powerful starting point for creating a playlist or as an organizing principle for a music collection. In every-day language, the terms music preference and musical taste are intuitively meaningful and apparently self-evident. They are interchangeably used to refer to the same concept. We make a distinction between the two, following the definitions as given by Abeles [1]. Musical taste is defined as a persons slowly evolving long-term commitment to a particular music idiom. Its development is assumed to depend on the cultural environment, the major Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. ' 2002 IRCAM Centre Pompidou

PATS: Realization and User Evaluation of an Automatic Playlist Generator consensus [3], peer approval, musical training [4], age as an indirect factor [5][11] and other personal characteristics. Personal music acquisition behavior over time is likely to represent the development of a persons musical taste. On the other hand, music preference is defined as a person s temporary liking of particular music content in a particular context-of-use. It is instantaneous in nature and subordinate to the musical taste of a person. Music is deemed to be preferred if its musical features suit particular activities, moods or listening purposes. Therefore, the context-of-use is supposed to produce constraints and opportunities for what music is preferred. It sets what kind of music should be selected and what kind of music should be rejected. North and Hargreaves [10] showed that music preference is associated with the listening environment and that people prefer to use different descriptors for music to be listened to in different environments. For instance, music for a dance party sets up desirable and undesirable criteria on tempo, rhythmic structure, musical instrumentation and performers, which are likely to be different for a romantic evening, for dull or repetitive activities or for car traveling. However, an indefinite number of contexts-of-use may exist they all produce different criteria for preferred music. In addition, the particular experience to listen to given music does not need to be the same in similar contexts-of-use or a given context-of-use is unlikely to be best provided with exactly the same music, over and over again. In other words, music preference changes over time. 2.2 Interactive control of PATS When using PATS, the link between a context-of-use and a playlist is established by choosing a single preferred song that is used to set up a complete playlist. Thus, music listeners only have to select a song that they currently want to listen to or that they prefer in the given context-of-use. This selection requires minimal cognitive effort as it may be the result of habitual behavior or affect referral. People may choose a song that is chosen always in a similar context-of-use, that was selected last time in a similar context-of-use, or that was given much thought lately. After selecting a song, PATS generates and presents a playlist, which includes the selected song and songs that are similar to the selected one. While listening, a music listener indicates what songs in the playlist do not fit the intended context-of-use. As only a decision of rejection is needed for a small number of songs, this task makes only a small demand on memory processes. This user feedback is used by PATS to learn about music preferences of the listener and to adapt its compilation strategy for future playlists. If the system adapts well to a listener s music preferences, user feedback is no longer required. Moreover, PATS does not require any other user control actions. 2.3 Requirements Ideally, PATS should make music choices that would have been made by the music listener in case no PATS was available. Therefore, it uses attribute information of music on which human choice is largely based, and generates playlists that are both coherent and varied. Jazz was chosen as a music domain in this long-term research project, as jazz contains a variety of well-defined styles or time periods serving a diverse listening audience and its appreciation is largely insensitive to temporarily prevailing music cultures and movements. 2.3.1 Attribute representation (meta-data) of music Music listeners use many different musical attributes for their music choice. Talking about and judging popular and jazz music in terms of musicians, instruments, and music styles is common. It is therefore reasonable to represent songs as a collection of attribute-value pairs (meta-data). We have created and collected an attribute representation for jazz music of 18 attributes, in total. Their values were primarily extracted from CD booklets, discographies, books on jazz music education and training, and systematic listening. A listing of all attributes and an instance is given in Table 1. Table 1. Attribute representation for jazz music. Title Title of the song All blues Main artist Leading performer/band Miles Davis Album Title of album Kind of blue Year Year of release 1959 Style Jazz style or era postbop Tempo Global tempo in bpm 144 Musicians List of musicians Miles Davis, John Coltrane, Cannonball Adderley, Bill Evans, Paul Chambers, Jimmy Cobb Instruments List of instruments trumpet, tenor saxophone, alto saxophone, piano, double bass, drums Ensemble strength No. musicians 6 Soloists Soloing musicians Miles Davis, John Coltrane, Cannonball Adderley, Bill Evans Composer Composer of the song Miles Davis Producer Producer of the song Teo Macero, Ray Moore Standard/Classic Standard or classic jazz song? Yes Place Recording place New York Live In front of a live audience? No Label Record company CBS Rhythm Rhythmic foundation 6/8 Progression Melodic/harmonic development modal Results of a focus group study showed that the set of attributes and their values is sufficient to express reported preferences for jazz music. In this study, participants were instructed to assort a set of 22 jazz songs into a preferred and rejected category and verbalize their decisions. Many of the criteria elicited could be expressed as a logical combination of attribute-value pairs. 2.3.2 Wish for coherence Coherence of a playlist refers to the degree of homogeneity of the music in a playlist and the extent to which individual songs are related to each other. It does not solely depend on some similarity between any two songs, but also depends on all other songs in a playlist and the conceptual description a music listener can give to the songs involved. Coherence may be based on a similarity between songs such as the sharing of relevant attribute values. When choosing music, music listeners tend to focus on relevant attribute values for reducing the available choice set of songs and for making different songs comparable. This includes eliminating songs with less relevant attributes values and retaining only the ones with the more