COIN Course 2009

Emerging social networks of unsigned country music artists on MySpace: Using demographic and network related information of an artist friend network Jonas Kuhlemanna, Matthias Rottländerb, Ville Heikkiläc [email protected] b [email protected] c [email protected]

Abstract

The Internet has become a large distribution channel for many industries. Even so, especially for the music industry. However, for record labels online social networks could allow to identify rising artist earlier. In this paper we discuss a data driven approach using friend network information of 100 unsigned country music artists on MySpace We measure the Kulback-Leibler-Divergence of geographic location based on friends attributes in order to predict the potential to emerge. In addition other demographic and network related characteristics were used for further investigations. We have found that the geographic location is a relevant characteristic of each network and therefore it can be assumed that the online adoption process of fans is not very different from the offline world. Keywords: Social Network, MysSpace, emerging music artists, Kullback-Leibler-Divergence

1. Introduction An unsigned less known artist usually gains popularity at her or his hometown before reaching wider audiences. In the beginning, unsigned artist‟s growth is strongly correlated to the word of mouth effect. The social network around the artists is an important factor for successful growth in the early stages of the artist‟s career. Early adopters spreading the word of mouth can be more effective than the artist, especially during a period when no label or investor normally recognizes the artist. Usually the word of mouth effect is much stronger in a network where actors are highly homogeneous. If a friend of an actor is a fan of an emerging artist, it is more likely that the friend becomes a fan too compared to a complete random actor somewhere the country. However, the location-based growth in one region will reach its maximum when the homogeneous network stops growing and structural holes to other actors appear. At this point, promotion and advertising are important for further growth. Successful advertising can provide the network with additional random ties outside the homogeneous clique and help the network keep growing. This location-based growth effect may be expected in an offline world where transaction costs are high and random ties are less likely. However, in the ages of the Internet and social Media, the transaction costs are only a fraction of past costs. Two questions then rise: Is the adoption process in internet driven social networks the same as in the offline world, or is it different? Is it possible to measure an emergent social network by analyzing demographic attributes and network related information of an artist‟s friend network? However, in this paper we discuss these questions by using a data driven approach using friend network information of unsigned country music artists on MySpace. After that, we describe a location-based approach by measuring the divergence of the actors‟ region within an artist‟s friend network. Then we measure the networks around several artists by looking at the network structure. MySpace is a social network site founded in 2003. The site is structured as a network of individual pages and links between the pages. Links are represented by friendship relationships, which are always bi-directional. Both parties have to accept the friendship invitation to be linked. MySpace was originally intended for private individuals, but the service has since expanded to cover other kind of profiles such as bands, interest groups and commercial actors. Since acquisition by Rupert Murdoch's News Corporation in 2005, additional services such as videos and browser games have been added to the site. In the beginning, the site offered users limited capabilities for customizing their page. Over time, the customization options have been expanded following the introduction of different www-technologies such as CSS, JavaScript and AJAX. As a result, pages in MySpace are built on many different www-technologies. Older pages may use only static HTML while recently added pages may contain AJAX-based dynamic functionality. MySpace users can edit parts of their page‟s HTML and CSS code by hand. This results often in pages that are not valid HTML or CSS, which complicates automatic analysis of page contents by e.g. screen readers or crawlers. Typical personal MySpace page contains at least the two following sections. General information of the owner of the page, such as user name, country, state, city, sex and age. Short list of top few friends and a link or an AJAX gadget that allows browsing the whole friend list of the owner. There is no concrete way to automatically differentiate between real personal pages and other kinds of pages, except for sub-domain music.myspace.com, which is reserved for musicians and bands. This makes mapping friend networks of real persons challenging. MySpace Music is a specialized part of the site designed to provide centralized location to find music, bands and artists in MySpace. The music site offers charts of many genres of music in different countries. In addition, it is possible to filter artists by whether the artist is unsigned, independent or signed to major label.

2. Methodology 2.1. Divergence measurement using location based attributes In the ages of ubiquitous online social networks an unsigned artists can create his own fan network easily. However, the size of the network is strong driven by the artist‟s popularity. This characteristic may help to identify popular 2

artist, but it is not suitable to predict an artist emergent. A static view on the absolute size of a fan network is representing the result of the successful adopting process in the past. A different perspective based on observation in the present is needed to predict that an artists‟ network will grow in the future. If we think of an artist that is first known at one location and less popular in other regions, it is to be assumed that when she or he is getting more popular, fans in other locations appear. Over the course of time the local rootedness will decrease and the popularity increase. This effect is shown in Figure 1.

Presumed ratio between divergence and growth 2

3,00

1,5

2,00

1 1,00

0,5 0

0,00 500 1000 2000 4000 6000 8000 10000 growth

Divergence Figure 1

Thus, we formed our research hypothesis as follows: The sudden increase of an artist's degree centrality in her or his friend network in the same geographical area where the artist resides predicts that the artist is going to emerge (in immediate future)

2.2. Kullback-Leibler-Divergence The idea to use Kullback-Leibler-Divergence is based on the paper: “Spotting Out emerging artists using geoaware analysis of p2p query strings“ [1].It is a non-symmetric measure of the difference between two probability distributions P and Q, where Q is like a model or approximation of P.

The authors suggest using the divergence in order to predict the possible success of new products assuming a location based adopting process. In case of geographic location, it is a measurement of comparing an observed location distribution in one region to a uniformed location distribution including all regions. Because the paper describes a data approach based on P2P query strings, we had to adopt this approach to measure the location rootedness of artist in online social networks. We used the region attribute of each fan respective actor within the observed population and associated each fan with an artists‟ network. The population of all regions during the observation was used as a approximate uniformed distribution Q. Thus, for our purposes P(i) and Q(i) are redefined as: 3

Every actor within an artist‟s network was seen as friend or fan of the artist The distribution of regions P(i) for each artist a was calculated by dividing the total number of friends observed in region i by total friends of artist a. Q of region i was calculated by total friends in region i divided by the number of total observed friends. Assumptions: It is more likely, that artists with many friends have low divergence value because their name recognition is higher and therefore they have friends in different regions, whereas an artist with only few friends has a high divergence. He is only known in one region, which usually is his hometown and he has only friends from this region. When an artist has almost more than 8000 friends the divergence is assumed to converge to zero. 2.3. The data approach We made the decision to analyze the web based social network MySpace for two reasons. First, we needed a social network that was big enough and where user data was accessible for us. On MySpace, user data and network information is more likely to be open to the public than in for example Facebook. Second, MySpace is highly focused on music related topics and has charts and rankings for unsigned artists. We decided to focus our observations to a specific music genre to avoid influences such as different trends in popularity and audience of a genre, and to assure that every artist can be measured by the same attributes. We chose the country music genre because we expected more homogeneous audience than for example pop music audience. Since MySpace has no official data-API support, we were forced to develop our own crawler. The crawler needed to be capable of parsing the DOM structure of a user profile page, the friend list and the quick info. Quick info pops up for each user on the friend list and is used by MySpace to support an internal Ajax function. Because the size of the HTML code of a typical user profile is about 100kb in average, we decided to focus on parsing the quick info page to increase crawling speed. The key technologies we used were PHP as a scripting language, including the web-query-library curl, and MySQL for data storage. Friendships between users were stored as links between actors and were read through the friend list page of each included user. We started our observations on December 18 th 2009 when we collected the first degree friend data (total 117.000 unique users) of the top 100 American unsigned country music artists from the MySpace chart list. For each user several demographic attributes were collected: friendID, username, region, city, age, gender, number of friends, and whether the user is an artist or not. See 1: User attributes for details. Attribute

Data type

friendID

unique ID (numeric)

Username

free text

Region

Text (ordinal) or free text

City

free text

Age

Numeric between 0 and 100

Gender

‘male’ or ‘female’

Number_of_friends

Numeric > 0

Is_band

Boolean Table 1: User attributes 4

To detecting network growth, we updated the network data for each artist on January 3rd 2010 and on January 11th 2010, and we added the user data of new friendships (2000 and 2000) created during this period. We also collected the second degree for chosen artists later on January 14th to get more information about the network structure.

Figure 2 : Data approach

2.4. Filtering and cleaning In order to get significant results it is necessary to identify actors who are providing value to the artist and to the network itself. The value of an actor is a matter of viewpoint, and thus it is a subjective process. For example, an emerging artist is most interested in expanding his or her friend network on a quantitative way to increase the level of awareness. Qualitatively, the artist prefers to have friends who have a high friend count in the network, or who otherwise increase attention. Few examples are celebrities, famous artists of the same genre, or talent and promotion agencies. The analysis process has a different perspective as it is not focusing on subjective interest in an artist but preferably gives an objective view over a set of different networks. Therefore, valuable actors are those who have an interest to unselfishly support the artist and therefore enrich the network with social value. Social value of a network describes a bulk of different network characteristics that enrich networks‟ power to grow and stabilize itself with the help of word of mouth and social communication. For example, celebrities provide quantitative value to the network but have a little to no effect on the overall social value. It seems necessary to filter out artist‟s friends who increase the quantitative value but provide little social value to the network. For example, celebrities, promotion profiles, politicians, companies and other users who do not belong to the real fan base. It is important that a deterministic process with clear rules execute such filtering so that the results are affected in a deterministic way and the impact on the analysis results is reversible and understandable. From the characteristics of the data, it becomes clear that an algorithmic approach of identifying actors by attributes or behavior is unfeasible. The only attribute that identifies an actor as a member of a specific group is „is_band‟ (see 1: User attributes). Hence, we needed to make a rough decision. We excluded from the crawling process all actors (except the observed artists) that were labeled as a band. This speeded up the crawling process and ensured that artists who tried to influence their quantitative growth by adding other famous artist as their friends did not influence the analysis. Within a small subset the attribute „Number of friends‟ was showing the strongest correlation with the social importance in an artist network. The frequency (see Figure 4) shows that different user patterns can be identified. Within the unfiltered result set, the average number of friends is 3528, which is not very representative due to a high dispersion of values. Several assumptions had to been made and were controlled within the sample space. A general assumption is that the more friends a user has the more likely she or he is not a representation of the real fan base. Users that have more than 50.000 friends can be clearly identified as people or organizations that are of general public interest 5

16000 14000 12000 10000 8000 6000 4000 2000 0 0 10 20 30 40 50 60 70 80 90 100

49000

43000

37000

31000

25000

19000

13000

1000 7000

400

freqeuncy

8000 7000 6000 5000 4000 3000 2000 1000 0 0

Number of Datasets

(Barack Obama, for example). A subset of actors that have between 5000 and 10000 friends are highly popular on MySpace but seem to belong to a group of people or organizations that are special interest groups such as country genre specific talent shows.

age

Number of Friends

Chart Figure 4: Frequency of1: number of friends

Figure 3 : Frequency of age

As we decided to filter out actors who seem to be less valuable, we deleted all actors (except the 100 observed artists) that had more than 5000 friends. The number of unique users was reduced by 4779(4%). Because some of the actors‟ attributes were freely editable text fields, the attributes were not covered by an endless homogeneous domain. By looking into a sample set, not all actors seem to have the same restrictions on editing text fields. We assume that MySpace made some changes in internal attribute design in the past and newer profiles have only the option to define their attributes in a specific domain, but older profiles were not affected. As MySpace does not check the validity of attributes, we had to accept that there is some probability of fake user profile data. In the case of age information, fake data can be identified by measuring frequency of several classes of age (see Figure 3). By looking at the symmetry, the age “0”, “10” and “100” show a lot more members than expected. Because actors with the age “11”, “12” and “13” did not exist within our datasets, it is very likely that this information provided by the user is fictitious. In addition, it can be assumed that the frequency of actors that are older than 90 is oversampled. This fact was leading us to exclude these actors from the average calculation. Domain all actors all actors location measurement age

Filter rules exclude: actors > 5000 friends exclude: bands exclude: regions that have only one member exclude: actors that are younger than 14 and older than 70

Effect delete 4779 unique user focus on real fans exclude fake regions exclude fake ages

Table 2 : filter rules

6

3. Results 3.1. Demographic analysis of emerging artists All 100 artists were country artists listed in the unsigned artist charts of MySpace, but their individual network characteristics showed a high variance. The number of fans of each artists who passed our filter criteria resembles a Powers distribution (see Figure 5). The top 23 artist have more than 80 % of the total fans in the top 100 charts and this indicates the existence of several different patterns of artists‟ friend network. First, artists who seem to be very popular in this specific genre. This suggests that these artists have probably already emerged, but for some reasons are still marked unsigned. Second, artists that have strong fan base but who are not popular amongst the general public. We assume that this pattern contains artist who have the potential to either emerge or stagnate in growth due to their existing fan base. Third, artist networks where the existing fan base is too small to be used as an indication of success. This so-called long tail segment showed the highest percentage of growth. (See Chart 2 : ) totalFriends

network growth (Dec 18 to Jan11) 60,00 %

growth

friends

30000 20000 10000

40,00 % 20,00 % 0,00 %

4 87 95 98 92 6 9 86 70 54

0 artist

Figure 5: Distribution of network size (Dec18)

1 8 15222936435057647178859299 artist

Figure 6: Distribution of network Chart 2 growth : *excluded artist 44

In general, the network growth showed a high variance through all the different networks. A remarkable relative growth showed artist 44 with 14050% growth over the observation period. He started with two friends and ends up with 283 friends. Measuring by his absolute network size he was located in the third segment discussed above, but we decided to include him in further investigations. Artist 62 showed the highest growth (11.75%) amongst artists that had more than 500 friends at the time of the first observation. Several artists showed a very limited relative growth. This case occurred in all three segments. As the reason was not obvious, a further investigation was necessary. Artist 89 showed the smallest relative growth (1 friend, started with 799) in the segment of artists with more than 500 friends.

3.1.1. None location attributes To gain first insights into the characteristics of our datasets, we explored two demographic attributes of each actor: gender and age. Most actors were female (57%). Only 0.5% of actors did not specify their gender. The distribution by artists showed a high variance (see Figure 6). Interestingly we found three networks extremely dominated by female actors (with over 90 % female), but in the inverse case only 70% were male. It seems that these artists were focusing on female fans or have some characteristics that especially attract women and reject men. Two of these female fan dominated networks (networks of artist 62 and 44) had also the second strongest absolute growth over the

7

observation period. Due to a weak correlation (0.23) between percentage of female fans and network growth, no statistically significant correlation between the ratio of females and growth could be made. Generally speaking the average age of the observed actors is 30.94 (calculated by including several filter rules, see Table 2 : filter rule). It is worth mentioning that there are several gender specific differences in the distribution of age. Women are mostly younger between 16 and 24 years old, while men dominate in the age group of 30 years and older. Plotting the frequency of age against gender also shows that either men tend more towards faking their age, or both age and gender are given fake values together. This can be shown based on the assumption of faked age attributes discussed in 2.4. Peaks located at the ages of 10 and 100 show that those ages were probably values faked by most the users of MySpace. However, based on our data, we cannot conclude that men are more tending to fake their attributes in general. Including the attribute number of friends additionally shows that male in average have more friends (795.66) than female (603.01). In summary, it can be stated that there are indeed gender differences in usage behavior.

16,00%

100%

14,00%

80%

12,00% 60%

10,00%

40%

8,00%

female %

6,00%

male%

20%

4,00%

0%

Male %

Female %

Unknown %

Figure 8 : Distribution of gender

0,00% 0 10 20 30 40 50 60 70 80 90 100

44 21 8 26 46 61 94 29 34 39 41 99 27

2,00%

Figure 7 : Distribution of age by gender

3.1.2. Geo-location The geographical location of actors was interesting in order to gain some insights into the general roots of American country music, but however, the individual distribution of a geo-location for each artist network relatively to all observed geo-location is much more powerful, because the location of a network is, as we will show later, an individual characteristic of each network. We used the region attribute extracted from the available actors and calculated an individual K-L-Divergence value for each of the 100 artists on each observation period by the model we introduced in 2.1. In general the model outputs values between 0 and 4. The first observation period allowed a static view on the distribution only, but helped us to draw conclusions on the general relevance of location. Figure 9 shows that artists beyond 5000 friends throughout have a low divergence value, which means that artists of this segment are not rooted to (only) one location. Whereas artists with less than 5000 friends more vary, the artists with a small fan base are more likely to be strong rooted to one region. Interestingly, several artists are strong rooted to one region and still had a strong fan base beyond 500 friends. It can be seen that the geo-location of the observed online social networks is not totally random distributed, which support our assumption that the adopting process in online social networks does not significantly differ from the offline adopting process. 8

artists K-L-Divergence

5,00000000 4,00000000 3,00000000 2,00000000 1,00000000 0,00000000 0

5000

10000

15000

20000

25000

fans(Dec 18)

Figure 9 : Distribution of K-L-Divergence

In order to prove or disprove our hypothesis a static view on location characteristics was not quite appropriate. Therefore, we included the attribute growth and a time dependent location divergence measurement into the analysis process. The divergence was calculated accumulatively, which means that the measurement included all actors existing in the database to the certain point in time. Growth was measured by the increase of new friendships of each artist within the two observed periods. Most networks showed a decreasing divergence, only a few increased their fan base in the originated regions. The correlation between the relative change of location divergence and the relative network growth was strong negative (-0,8468), what indicates the fact that the greater the network growth the more a network is tending to loose the regional roots. De facto, this observation was influenced by a high variance and therefore it has to be said, that this cannot prove or disprove our hypothesis. So the correlations between growth and the shift of location as an indicator of emergence was still unclear. But this step was helped to indentify several patterns. Artists with an initial big fan network tend to be more regional scattered compared to some smaller artists. A general trend for this segment of artists cannot be seen within our observation, because some of them increased their regional roots, some were not. Artist 62 was the only artist that had both, a remarkable absolute growth of 313 fans, which is the second highest value within the total observation period, and an increasing divergence value. Medium-sized networks (800 to 2000 friends) are more likely to have strong regional roots at one location, but due to a high variance, this cannot be generalized. Artist 89 for example had strong regional roots and at least 800 friends, but his growth seemed to be stagnating. Whereas artist 90 (660 friends) with a low location divergence was still growing at different regions. This indicates that strong roots at one location for medium-sized artists can have a growth limiting effect as well. Assuming that these limited growth is caused by a too homogenous network, it is imaginable that an adopting process that serves the network with new actors from different regions can break through stagnation. Thinking of advertising or promoting an artist, like a music company does, would probably be an appropriate time to start. Because of the massive growth of artist 44 in the observation period, it was interesting to see how his location divergence was affected by that. His divergence decreased by 88 % from 2.24 to 0.25. It should be noted that this high divergence value on December 18th was less significant, as the network had only two actors at that time. But comparing his divergence on January 11 with similar artists, his divergence was still far below average. This indicates that his growth was caused by new fans that came from very different regions; therefore he did not show the location based adopting effect.

9

3.2. Network analysis The attribute based analysis helped us to understand how location was distributed but could not reflect significantly how the network growth was effected by the “Word-of-Mouth-Effect”. Interconnectedness between fans of the same artist was not observed yet. Therefore, we further decided to investigate the fan network of chosen artists of the observed different patterns: Artist 1, Artist 44, Artist 89 and Artist 89 3.2.1. Interconnectedness and growth Artist 1 During our observation period, artist 1 dominated the unsigned country artist charts on MySpace. Although we found out that the artist has a record deal with Parallel Entertainment, was listed in the unsigned charts. Regarding to his network size (8011 friends) this seemed to be a network that was already emerged. His location divergence (0.1129) is low compared to other observed artists. By looking at the network Figure 11 shows the friend-network from artist 1. The artists and those friends who are exclusively connected to him are disregarded. The network possesses a high density and a lot of friends coming from the artists‟ native region. Regarding its size, this network showed characteristics we expected to find. We assumed that this artist has already emerged, because he has already released his first album. We also measured the centrality from the network shown in figure 10. Betweeneess centrality is a measure for the relevance of one actor to the network [2]. To be clear: betweeneess is a measure for the actor„s impact on the communication of other actors. An actor is called central when he is the connecting link of other relationships. As shown in figure 10 Eric Wayne has the highest betweeness value in artist 1‟s network. Further investigations showed, that Eric Wayne is a radio host at Big Country 99.5 in Oklahoma; a broadcast station, which airs mainly country music. If Eric Wayne likes the music of artist 1 his songs will probably get extensive airplay on his radio show. Therefore he functions as a promoter. Because of this artist 1 has a good chance of emerging even further. In this case the betweeneess centrality helped us to indentify a valuable person, who provides social value to the network. But this was only applicable due to our filter approach.

Figure 11 Artist 1 // friends: colored by region // artist: excluded// communication frequency:2

Figure 10: Artist 1 // friends: colored by betweenes centrality / artist:excluded// communication frequency:2 // b 10

Artist 44 At the beginning of our investigation the Artist 44 had only two friends. At the end of the observation the Artist 44 has already 283 friends. The low divergence in spite of the relative small number of friends is remarkable. In Figure 12 the artist‟s friend-network is shown. The friends are coloured by region and surround the artist positioned in the centre of the graph. A very low degree of interconnectedness between his friends on the one hand but a high degree of friends from different regions on the other hand can be observable. Because of this data, can be assumed that the friends of the artist are randomly generated by sending friend requests by the artist. Checking the comments on the artist‟s profile there are lot of acknowledgements for been added. This is also a hint that the artist himself is the prime mover of his growing network. He has only a few friends stemming from his native region. Although the internet theoretically overcomes regional boundaries our data hint at similarity between the network growth “onand offline”. In our opinion, the network growth is not driven by internal factors like the “Word-Of-Mouth-Effect” because of low network density. Therefore the artist 44 has not a good chance to emerge.

Figure 12: Artist 44 // friends: colored by region // artist: included in center

Artist 89 The artist 89 starts with 799 friends. At the end of our observation period, he has gain only one new friend, so stagnation has set in. The divergence has a high degree (2.43) that means he has many friends in one region. 2/3 of his friends stems from his home region.

11

Figure 13: Artist 89 // friends: colored by region // artist: not included // communication frequency: 2

Figure 13 shows only the members of the network, which are interconnected. Friends only connected to the artist, are excluded. In the red marked array the phenomenon of “word of mouth (WOM)” is observable. WOM characterizes the dissemination of information within a community. Actor 1 advises actor 2 the music of artist 89. Either actor 2 likes the music or he does not. In this case actor 1 has got three actors interested in the music of artist 89. The network is built up by actors, stemming from four different regions. The color grey represents Illinois whereas magenta represents Indiana. These two regions are geographically neighbored. Maybe this fact facilitates the spreading. The artist seems to have become popular in his home region but his name recognition stagnates. Due to the local demand a quick national advertising campaign is suggested.

Artist 62 During the observation period Artist 62 has the highest absolute growth (360 additional friends) by all examined artist. Figure 4 shows only the members of the network, which are interconnected. Friends only connected to the artist are excluded. The artist has an increase in divergence whereas he has a strong growth. The growth concerned only one region. In the highlighted array in Figure 14, a clique is observable. Within graph theory a clique is defined as a group of at least three actors directly linked to each other [3]. Since cliques fulfilling this condition rarely contain more than three or four actors the concept was altered into the so called n-clique-concept: Instead of the criterion of each member‟s direct linking it is postulated that every member has to reach other members with relative short npaths. E.g. a 1-clique is equal to the strict definition of cliques (n=1; i.e. direct connection between each member). A 2-clique (n=2) meets the condition that every member has to be reachable by two paths and so forth. The clique observed in artist‟s 62 network can be characterized as n>1 clique. It can be assumed, that the artist‟s music is the shared interest of this clique. The high interconnectedness within this network implies a good chance to become successful.

12

Figure 14: Artist 62 // friends: colored by region // artist: included in center // communication frequency: 2

4. Conclusion 4.1.1. Lessons learned

All in all, our investigation of different artist„s friend-networks on MySpace shows that location is a relevant characteristic of emerging artists‟ network. The distribution of region within the observed networks cannot be seen as a random distribution. We assumed that an artist will have a good chance to emerge if the regional network grows and the divergence increases. This development was observed for some artists but the data are insufficient to make a significant point. It is to be assumed that other parameters affect the development which could not be considered within the scope of our investigation. Therefore the correlation between growth and local rootedness could not be proved. In the “offline world” an unknown artist first becomes known in his native region and then his name recognition could increase national by targeted advertising. Due to the relevance of location in the observed online social networks it can be assumed that the online adoption process of fans is not very different from the offline world. For record companies the information of when to start to promote an emerging artists online is essential. Our findings that some artists are popular and also strong rooted to one region, but the network was stagnating in growth maybe help record companies to identify the targeted segment and the right point in time to start with promotion. When the growth of the friend-network remains static for the native region of the artist it has to be assumed that the region is saturated. For emerging purposes the artist should be increasingly advertised and live shows in neighbouring regions should be offered at that time. Boosted by local support his level of familiarity grows also in other regions. On the basis of our findings gender-specific usage behaviour can be recognized. There are some networks of artists which are predominantly joined by women (90%) whereas men are hardly represented (≈ 10%). On the opposite the biggest male-dominated networks have a higher percentage of women (30%). Men rather tend to name the wrong age. Profiles presenting “10” as age statement more often stem from male participants. The growth of friend-network depends on many parameters which cannot be exhaustively measured. Artist 44 for example has many friends from different regions and only a few friends stemming from his native region, but was growing very fast during the observation. The density of the network is low. Hence the social linking within his network is also small. This type of growth independent from the geographical location is uncommon compared to

13

other observed country music artists. Due to very low interconnectedness of fans it is obvious that his network growth was not driven by network internal factors such as the word of mouth effect.

4.1.2. Limitation of the research As the research was more or less an experiment in a semi structured setting it was hard to prove or disprove our hypotheses, but it opened up interesting insights in some outstanding patterns. Several uncertain factors have an enormous impact on our results. First of all, the observation period possibly influences the acquired data and therefore our conclusions. It is unclear, which impact seasonal fluctuations have on user activity, growth and trends. In terms that our observation was taking place during the Christmas days, it is still unclear if this circumstance was increasing or decreasing user activity or had no influence at all. So the question about the significance of the data acquired in the limited time period should be made. The emerging of an artist is a process which takes a lot of time. To draw conclusions on that it is necessary to observe several influences that where caused by past issues as well. This cannot be granted within this small period of observation. Due to several semi structured issues, we were forced into a rough filter approach, which had an impact on our results (see 2.4). The huge amount of data could not be handled manually, so the filtering process had to be simple and therefore automatic. The circumstance that an artist can invite friends by himself and mostly less popular artist have the affinity to increase their network growth quantitatively, induces a uncertainness about the real value of network growth. Also some attributes were not defined within a structured domain and therefore they could not be used in an deterministic automatic approach. This forced us to exclude those from the analysis process. In measuring the interconnectedness between fans, we had to access a list of friends of each user. These information were only assessable by one out of two, because 50 % of the observed users do not want to present their friend information to public. Because friendships are bi-directional, the chance to find an existing friendship within an artist network is still greater or equal to 50%, but there is still a subset of friendships within the network that cannot be identified. Because of poor data quality, the region attribute has to be used instead of city, to define the geo-location of each actor. The region covers a fairly wide distance, so that the allocation of each actor to a local subgroup is very rough. Overall, we had to accept a certain fuzziness, which had an impact on our research process and therefore on our results. 4.1.3. Future research To reduce possible influences caused by a limited time horizon the observation period should be extended in future research. Six or more month should be adequate. This will not only reducing the uncertainness of seasonal fluctuations but increasing the probability to identify phenomena that have an influence on an artist emergent. As rough filtering influences the accuracy of possible conclusions that follows from the data, a finer filtering process would lead to better results. By that it makes sense to focus on developing a more complex method or algorithm that filter out less valuable actors more precisely. Also a more flexible method of sorting and identifying information within free text fields, especially the city and region attribute, would help to evaluate the given data more precisely and therefore significantly. More demographic attributes, such as hobbies or music taste combined with geo-location attributes will gain insights into how homophile of actors from the same region affects the power of network growth. Because our observation showed that the distribution of geo-location is less likely to random it is imaginable that other demographics will show the same. Combining the geo-location with demographic attributes, which do not follow a uniform distribution, could help to indentify finer patterns of networks and therefore find characteristic or phenomena within different phases of an artist emergent. When you figure an artist as a product, it is imaginable to transfer this approach to none music related topics, such as product marketing.

14

5. Literature [1] [2] [3]

Noam Koenigstein, Yuval Shavitt, Tomer Tankel: Spotting Out Emerging Artists Using Geo-Aware Analysis of P2P Query Strings. KDD‟08, August 24–27, 2008, Las Vegas, Nevada, USA. Dorothea Jansen: Einführung in die Netzwerkanalyse Grundlage, Methoden und Forschungsbeispiele 3., überarbeitete Auflage, August 2006 Stanley Wassermann, Katherine Faust: Social Network Analysis Methods and Applications, 2009, USA

15