Music Intelligent Agents: A business model?

Table of Contents

Table of Abbreviations

1 Introduction
1.1 Basic problem and objective
1.2 Methodology and outline

2 Basic concepts and definitions
2.1 Intelligent agents
2.2 Recommender systems and information filtering

3 A Business Model for Music Intelligent Agents
3.1 Value proposition
3.1.1 Inconsistencies in the music market
3.1.2 Overview of the solution and benefit
3.2 Architecture of value creation
3.2.1 Creating user models
3.2.2 Recommendation interface and learning from user feedback
3.2.3 Filtering and recommendation technique
3.2.4 Software implementation
3.3 Revenue sources
3.3.1 Content provider
3.3.2 Consumer
3.3.3 Market research data and targeted marketing

4 Further implications of the model
4.1 Technical limitations and challenges
4.2 Security and social issues

5 Conclusion

Appendix

References

Table of Abbreviations

illustration not visible in th

1 Introduction

1.1 Basic problem and objective

The rise of the Internet brought disintermediation effects to a huge number of industries of which the music industry is maybe the most prominent example. How technology changed the rules of this world-wide multi-billion dollar business became particularly evident mostly because of the important legal aspects involved.

Record companies had to put up with huge losses, and now on a sudden four years are gone since Napster sparked the P2P revolution in 1999.^[1] The increasing digitization of music and concomitant changes in music production and distribution are not reversible anymore. Astonishingly, record companies have not yet come up with models or strategies that use the advantages of this process successfully. Any attempt failed so far because they could not convince the most important stakeholder – the consumer.^[2]

This paper concentrates on intelligent agent technology which has been promisingly described as “[..] software that is proactive, personalized, and adapted. Software that can actually act on behalf of people, take initiative, make suggestions [..]”^[3]. With more than 4 million existing music titles^[4] music listeners have increasingly problems to discover and select artists or songs they really like. Consequently, they tend to take (mostly in MP3 format) what they can get, be it legal or not, and hope to find a ‘jewel’ among the songs. Intelligent agents in the form of recommender systems pick up the idea of personalization which lacks application in the area of music compared to other industries and might help to mitigate the dilemma of the music industry.

Therefore the objective of this paper is to develop a business model based on intelligent agent architecture to automatically provide and recommend music that meets the individual consumer’s taste addressing both business and technical aspects.

1.2 Methodology and outline

This paper is structured as follows. Chapter two focuses on technical aspects and describes the most important definitions and concepts, and thus constitutes the theoretic basis for the subsequent chapters. It begins by explaining intelligent agents in a general context. Then recommender systems, an application field of intelligent agents, are explained. In order to describe recommender systems properly, it is necessary to address and explain the issue of information filtering. Finally, two information filtering techniques that are frequently used in recommender systems are briefly discussed.

In chapter three, which constitutes the main part of the paper, a business model for a music recommender system based on intelligent agent technology is developed. The business model is subdivided into three parts. First, the model is introduced and it is explained which benefit could be reaped by the various parties. Second, by defining the architecture of the system the process of value creation is described and the implications for directly involved stakeholders are explained. Third, potential revenue sources are analysed and involved business aspects examined.

Chapter four explains further implications of recommender systems. A selection of technical limitations and challenges is described and related to the previously developed business model. Additionally, security and social issues that are critical for the success of the model are discussed.

Chapter five summarizes the important findings of the study and closes with concluding thoughts.

2 Basic concepts and definitions

2.1 Intelligent agents

Intelligent agents (IA) also referred to as software agents and bots^[5], are electronic pendants to human agents in form of software programs. Just like travel agents or any other person that is commissioned by a customer, IA autonomously execute tasks on behalf of another party in a computing environment. These tasks usually require specialist knowledge and consist of a time consuming process.^[6] In the networked information age IA are increasingly applied to mitigate the problem of information overload and hence, they are also strongly related to the fields of information filtering and data mining.^[7]

Already in 1995 assessed as one of “the fastest growing areas of research and new application development on the Internet”^[8], IA today are vastly commercially used and help supply and demand get together more efficiently. In the B2C sector they e.g. support the consumer’s purchasing decision by comparing characteristics and prices of products (Mysimon.com; Compare.net; Edgegain.com), or as a recommender system IA suggest a product that ought to meet the consumer’s preferences (Findgift.com; Amazon.com). In the field of B2B, e.g. with procurement, IA interact and negotiate autonomously with each other on behalf of their ‘client’-companies.^[9]

The exact definition, however, of what can be named an ‘intelligent agent’ is not clear among users and researchers and there exist various types of applications which claim to incorporate IA technology. Especially the prefix ‘intelligent’ evokes controversy since it is the decisive factor that distinguishes IA from ‘normal’ software programs and search engines.^[10] In addition to the ‘seek and match’ capabilities of search engines, IA can monitor and proactively take action on movements on a Website or any other environment without the intervention of the user. So called learning agents are able to learn the individuals’ preferences and to make suggestions.^[11] According to Nissen, the element of ‘intelligence’ of an agent is the ability for adaptive reasoning which implies the “[..] capability to process information from external environments -- such as networks, databases, and the Internet -- given a set of knowledge, attitudes, and beliefs of the user which are understood by the agent”.^[12] One successful type of application of IA are the so called recommender systems which will be presented in the next section.

2.2 Recommender systems and information filtering

Recommender systems (RS) employ information filtering algorithms in order to provide product recommendations or suggest objects of interest to users.^[13] They work on the natural principle of relying on other people’s recommendations (e.g. book and movie reviews) when a choice needs to be made. Hence, RS typically collect information or recommendations from people and then process and redirect relevant suggestions to suitable recipients.^[14]

Information filtering is the process of seeking, selecting and delivering information to users. The goal is to provide users with their specifically required information that needs to be selected from an often large and dynamically generated amount of data. Therefore information filtering systems have a time saving effect and are indispensable in the information age.^[15] Different information filtering approaches have emerged, of which ‘content-based filtering’ and ‘collaborative filtering’ present the most renowned and discussed ones in the context of recommender systems.^[16]

The content-based or cognitive^[17] filtering approach is generally used to analyse and find items by means of keywords or text (content). Items such as newspaper articles can be easily identified by comparing keywords in the article with the user’s specifications. If, however, as opposed to an article, the searched item is music, video or of physical constitution, the system can only seek and find items when they have been assigned descriptions or attributes by hand beforehand. Pure automatic information filtering would not be possible.^[18]

Collaborative filtering (CF), also called social filtering, overcomes the described limitation. Consequently, it is suitable for music and other media. It is strongly connected and applied with RS, so that some authors favour using directly the term RS instead, although RS do not necessarily make use of the CF technique.^[19] With CF products are recommended to a new user based on the opinions of other users who like similar products as the new user.^[20]

3 A Business Model for Music Intelligent Agents

3.1 Value proposition

In the following the Music Intelligent Agent model is introduced. First, the problem to be tackled by MIA is pointed out. Then, the benefits of the model for the main parties of the value chain (content providers, consumers) are described.

3.1.1 Inconsistencies in the music market

Nowadays the music market is challenged by bringing supply and demand effectively together. Certainly the digitization of music made the distribution of music easier, but with the mass of digital content comes the selection problem for the consumer. A virtual music store today offers a selection that is maybe 1,000 times greater than an average conventional music store. But how does a consumer find a song he really likes out of 4 million?^[21]

Furthermore, the selection problem might also contribute to the habit of acquiring vast amounts of MP3s on the computer in an illegal way. If the consumer makes a couple of bad purchases because there was not enough assistance, the consumer might get irritated and there is a bigger tendency that he prefers to copy songs from friends. Consequently, the worse the matching between supply (song) and demand (consumer), the more people get discouraged to buy songs. The more songs are obtained criminally the more stakeholders (record companies, musicians, and finally even consumers) are harmed.

3.1.2 Overview of the solution and benefit

The proposed business model Music Intelligent Agent (MIA) is a RS based on intelligent agent software. MIA is able to learn the consumer’s preferences and to recommend music titles that match the taste of the consumer. As a result it customizes offerings to each single consumer which is also known as personalization - products are matched to individuals.^[22] Once MIA is ‘configured’ and running, recommendations are economically and automatically produced. In the value chain of music distribution, MIA is positioned as an intermediary between content providers (record companies and retailers) on the one hand and the consumer on the other and enables information exchange in both directions.^[23]

The consumer is given useful recommendations that might help him to find new songs or artists he can enjoy all life long. The effort needed is reduced to a minimum. In fact, his purchasing decision making is supported and his search costs reduced considerably.^[24] The benefit for content providers is convincing, too. MIA functions as a stimulus for the consumer’s need recognition by notifying the consumer of suitable products.^[25] It supports the content providers’ marketing efforts because more consumers get to know about more ‘suitable’ songs in reduced time which can result in rising sales. MIA is able to turn browsers into buyers and to enhance customer loyalty. Furthermore, gathered consumer data can be used for market research and targeted marketing.^[26]

3.2 Architecture of value creation

This section deals with the process of value creation. The involved steps are examined and the different stakeholders’ roles are explained.

3.2.1 Creating user models

Knowledge about the user is the starting point for MIA. Only after it has collected some data about the user’s preferences it can provide useful recommendations. Hence, creating expressive and exact user models is most crucial for MIA to be successful. A user model in MIA can be described as an interest profile and contains per song: the song title, the name of the artist and a rating, in terms of how much the user likes the song. There are different methods of gathering user data that can be distinguished by the degree of user involvement and by the type of input.^[27] In order to achieve best performance, MIA combines three different methods.

Exploring the user’s hard disk to find existing MP3 music files is a very implicit type of collecting the user’s preferences because the involvement of the user in terms of work equals zero. Although this is very convenient for the user and can be done in the background, the data is not entirely convincing. Owning a music title does not necessarily mean that the user likes it or would give it a high positive rating.^[28] Therefore, MIA uses this method as a basis for the next step of user modeling.

Out of the existing MP3 files, MIA chooses a minimum of x songs and asks the user to rate each of them on a scale from one to seven.^[29] Here the user involvement is high as he tells the system explicitly how much he likes the songs. In order to make this task easier for the user, MIA plays each of the songs in the user’s music player (e.g. Mediaplayer, Winamp) while the user carries out the rating.

In a third step, which is rather a continuous process, MIA analyzes and monitors what songs the user plays in his music player. Through this, songs are rated automatically by the system without a perceptible involvement of the user. This includes the songs the user has not rated manually so that finally any song the user has listened to at least once and those songs that have been rated manually, have been evaluated. MIA not only creates a list of what songs have been played how often, but also makes a note, when a song has been skipped. Field, Hartel & Mooij found out empirically that “people do not skip music they enjoy”^[30]. Accordingly, songs that are played frequently can go up in their rating whereas songs that are played seldom and / or are often skipped go down. In order to have an ongoing process of user modeling, the described methods are applied regularly and should always be accessible for the user. By that, MIA meets the demand of keeping user models up to date what is part of the next chapter.

[...]

^[1] See Brislen (2003); there is controversy to what degree digital piracy is responsible for decline in sales.

^[2] See Rojas (2003), the computer manufacturer Apple recently came up with an internet music platform (iTunes Music Store) that is accepted and successful.

^[3] Maes (1997), p. 11.

^[4] See Pachet / Roy / Cazaly (2000), p. 44.

^[5] See Patton (1999).

^[6] See Brenner / Zarnekow / Wittig (1998), p. 22; Nissen (1995).

^[7] See Turban et al. (2002), pp. 153-154.

^[8] Nissen (1995).

^[9] See Turban et al. (2002), p. 94 & pp. 156-158 & p. 250.

^[10] See Brenner / Zarnekow / Wittig (1998), p. 21; Turban et al. (2002), pp. 154-155; .Nissen (1995) is even more precise, he differentiates between programs, (software) agents and intelligent agents.

^[11] See Turban et al. (2002), pp. 154-160; Nissen (1995); Brenner / Zarnekow / Wittig (1998), pp. 25-29; Maes (1997), p. 11 & p. 17.

^[12] Nissen (1995).

^[13] See Mukherjee (2001), p. 114; Sarwar et al. (2000) p.158.

^[14] See Resnick / Varian (1997), p. 56.

^[15] See Baudisch (2001), p. 3.

^[16] See Chai / Vercoe (2000); Baudisch (2001), p. 7; Shardanand / Maes (1995).

^[17] See Olsson (1998), p. 6.

^[18] See Shardanand / Maes (1995).

^[19] See Resnick / Varian (1997), p. 56; Baudisch (2001), p. 12; Olsson (1998), pp. 6-7; Shardanand / Maes (1995).

^[20] See Sarwar et al. (2000), p. 160; chapter 3.2.3. explains how CF works in detail.

^[21] See Pachet / Roy / Cazaly (2000), p. 44.

^[22] See Turban et al. (2002), p. 133.

^[23] See figure in appendix p. 17.

^[24] See Turban et al. (2002), p. 61.

^[25] See Turban et al. (2002), pp. 125-126.

^[26] See Schafer / Konstan / Riedl (1999), p. 158; the benefits are discussed in greater detail in chapter 3.3.

^[27] See Claypool et al. (2001), p. 35.

^[28] See Schafer / Konstan / Riedl (1999), p. 164.

^[29] For the rating scale see appendix p. 17; see chapter 4.2 for implications of rating interfaces.

^[30] Field / Hartel / Mooij (2001), p. 7.