Thanks for the help:

Thanks for the help:

The effect of content-based recommendations on online display advertising.

By Joost Dorgelo

1

Thanks for the help:

The effect of content-based recommendations on online display advertising.

Master thesis

Faculty of Economics and Business Department of Marketing June 7th, 2017 University of Groningen By Joost Dorgelo Uithof 13 7761 XN Schoonebeek 06-10159446 j.dorgelo@student.rug.nl Student number: s3014649 Supervisor: prof. dr. J.E. Wieringa

Acknowledgment: I would like to thank Jaap Wieringa, for his support, insights and the way he motivated me to ‘go the extra mile’. I would like to thank Rik Smit, for the opportunity to

2

Abstract

Previous research on the effectiveness of recommendation systems, tested the effect of recommendations on sales. This research is performed to determine if recommendation systems also increases revenue, when a website is offering free content. It is believed that recommendation systems have this effect, because they increase the amount of visits and increase the click trough rate.. In this research, content-based recommendations, based on music type are used. The effectiveness of this type of recommendations is tested and

differences between music preferences, on the use of recommendations are investigated. The proposed type of recommendations did not lead to statistical significant evidence, to conclude that recommendations increase advertisement revenues. Also no evidence was found, that the recommendations based on music type are more effective than generic recommendations. A preference for Pop music is increasing the probability of the use of the proposed

recommendations system significantly. This leads to the conclusion that content-based music recommendations do not affect the revenue generated by advertisements. The implication would be that the implementation of content-based recommendations based on music types should be discouraged. This research has several limitations that weaken the strength of the conclusions. The most important limitations are the short observation window and the small set of recommendations to base these conclusions on.

3

Table of content

Motivations for using technology ... 9

Online recommendations ... 10

What music are people listing to ... 12

Music preferences ... 13

Differences in the use of music recommender system ... 15

Online advertising ... 16

3. Method ... 19

Recommendations based on music type ... 19

Finding similar users ... 21

Recommender adoption ... 22

Individual differences between clusters ... 23

Effects of visit duration and recommender use on click trough rate ... 24

4. Results ... 24

Classification ... 24

Clustering ... 25

Summary of the collected data ... 26

Effects of music type based recommendations ... 27

Amount of visits ... 27

Effect of different music preferences ... 27

Increase in click trough rates ... 27

5. Discussion ... 28

5

1. Introduction

“Music has always played a major role in human entertainment” (Kaminskas & Ricci, 2012, p.90.) but the rise of high speed internet made it possible to consume music differently (Fan, Kumar & Whinston, 2007). Before, records were sold on physical storage devices such as vinyl or CD, now most of the records are sold digitally via the internet. According to the International Federation of the Phonographic Industry (IFPI) the total revenue of the

recording industry in 2016 consisted for 50% out of digital revenues against 34% of physical record sales. The revenue from physical records was 7,6% lower than the year before. The rest of the revenue was generated by performance rights and synchronisations (IFPI, 2017). The increased availability of music on the internet has allowed for the development of a new business model, called streaming services.

Streaming services are granting the user access to their online music library. The popularity of these services is reflected in the digital revenues where a majority of 59% of the digital revenues come from streaming services (IFPI, 2017). There are paid and unpaid streaming services. The most popular paid version is Spotify, with fifty million paying subscribers worldwide (Statista, 2017). The video sharing platform YouTube is the most popular unpaid music streaming service. Worldwide 820 million people use YouTube to listen to music, making it the largest music streaming platform in the world (Statista, 2016). This makes watching/listening to music videos one of the core activities people visit YouTube for (Shao. 2008). This is also reflected in the top ten of most subscribed YouTube channels of February 2017, where five out of ten channels is dedicated to a musician (Statista, 2017).

The business model of paid services is charging a monetary fee to allow access to the

platform. With unpaid services the business model is often unclear and money is earned in an indirect way. Anderson (2006) describes that a common business model for free products is selling advertisement space. Users are granted access to the product at the cost of being targeted by advertisers. An example of a streaming platform that is based on advertisement is YouTube. A different type of business model is offering limited access to the product

features. Users can unlock full access by paying a fee. Spotify’s free version is an example of this model. The reason why free business models are attractive, is because they create the biggest reach for the product, creating a network effect where the platform with the largest

6 Because free music streaming platforms such as YouTube are dependent on advertisements, making advertisements more effective is essential to increase revenues for these types of businesses. The most important form of online advertisement is display advertisement, also referred to as banner advertisement (Ghosh & Bhantnagar, 2013). This type of advertisement is responsible for a large part of the revenue in the online advertising industry (Goldstein, Suri, Mcaffe & Ekstrand-Abueg, 2014 ). There are three parties involved with online display advertisement: the advertiser that wants to deliver a message; a publisher that allows

advertisers to deliver their message next to the content of the website; the website visitor that encounters the advertisement when visiting the website (Goldstein, Suri, Mcaffe & 2014 ).

The effectiveness of online advertisement is often questioned (Manchanda, Dubé, Goh & Chintagunta, 2006) and there is especially scepticism towards banner advertisement (Gosh & Bhatnagar, 2013). This scepticism is likely to be caused by the low click trough rates they generate, which is a common used metric for evaluating banner effectiveness. The click trough rate refers to the number of times people have clicked on banners, divided by the amount of visitors (Manchanda, Dubé, Goh & Chintagunta, 2006). Besides questionable effectiveness, another problem with banner advertisement is the increased avoidance of online advertising by visitors (Cho, 2004). Therefore the main challenges for advertisers and

publishers are improving the effectiveness of banner advertisement and try to tackle advertisement avoidance.

A popular way of increasing effectiveness is making advertisements more personal for the audience (Finn & Wadhwa, 2014). This is done based on data about the users’ online

behaviour. These types of improvements are mostly done by the advertiser themselves and are independent from the publishers. In order for a free music platform to be profitable (sell enough advertisement space), the presented content needs to be valuable for the user (Fan, Kumar, Andrew & Whinston, 2007). What is regarded as valuable content in the context of music, can differ very strongly across individuals and can be influenced by things like

personality or context where it is listened in (Rentfrow & Gosling, 2003; Kaminskas & Ricci, 2011).

7 industry is traditionally a blockbuster industry, meaning that a few players are dominating the industry with hit products (Anderson, 2006). The development from physical to digital goes hand in hand with a shift from blockbuster industry to an industry with increasingly more niches. This transition is believed to be caused by people looking for products that are more adapted to their own needs (Fleder & Hosanagar, 2009).

The availability of such a large amount of music has made it more difficult for people to search for music that is interesting to them (Kaminskas & Ricci, 2012; Chung, Rust & Wedel, 2007). The large amount of products available online, has caused that an increasing number of companies are trying to help consumers to decide by means of recommendations (Hennig-Thurau, Marchand & Marx, 2012; Fleder & Hosanagar, 2009; Kaminskas & Ricci, 2012). These recommendations are produced by algorithms fed with consumer data, called recommender systems. Most of these systems are based on the ratings that people give towards a specific item (Adomivicius & Tuzhilin, 2005), for example stars on a review website. However, only a minority of people are reviewing items. In 2015, only 10% of the people stated that they were systematically reviewing products and services online (PEW research centre, 2016). So often the rating for a product is not known and preferences needs to be derived from other types of data, for example historical data about the user.

Several studies have investigated the effects of helping the consumer with product choice. Häubl and Trifts (2000) determined that online decision aids decrease the amount of products that consumers are taking into consideration, but increase the quality of the products that are considered. In research done by Häuble and Murray (2003), it was found that products that are encountered by a consumer via a recommendation system, have a more prominent role in the decision making process of the consumer. Fleder & Hosanagar (2009) concluded that the use of online recommendation systems, increase sales. From these researches it is clear that recommendations have a direct effect on revenue and an indirect effect by increasing the quality of the purchase. However, recommendation systems are not only implemented at webshops, but also at website’s that depend on advertisements. An example of this are free news websites that recommend additional articles that might interest a user. It remains unclear if recommendation systems have a positive effect on revenue streams that are based on

8 that determine the revenue of online advertisement are amount of impression and the click trough rate (Chen & Stallaert, 2014; Ghosh & Bhatnagar; 2013; Manchanda, Dubé, Goh & Chintagunta, 2006).

This research will start investigating the added value of recommendation systems for websites that depend on advertisements. This is done by investigating the effects of content-based recommendations on the free online music streaming service Songa. The research question of this research is: what are the effects of content-based recommendation system on the amount of banner impressions and the click trough rate? The recommendations that are made are based on the music preference of the user. This music preference is determined by analysing what the users listened to in the past.

With this research, the knowledge on the effects of recommendation systems is extended to free advertisement based business models. This extension is currently not present in scientific literature. The results of this research could help in identifying drivers of successful online advertisement. The findings could help with the development of music recommendation systems, because the adoption of recommendations, based on music type will be determined.

From a managerial perspective, this research supports decision making related to the development of recommender systems. This support can help in determining if the development of a recommender system will generate enough extra revenue to cover its development costs. Given that there are concerns related to the decreasing effectiveness of online advertising (Ghose & Todri-Adampoulos, 2015), this research will help in finding new approaches to boost revenue through banner advertisement.

9 alphabetic order, followed by an appendix where several parts of the research refer to for extra (visual) clarification.

2. Literature review

This research starts with a review of the literature on what drives people to use technology. It is important to start with this topic, because this is the basis for the use of the recommendation system. The second part is about online recommendations. First the need for

recommendations is explained, followed by an explanation about how a recommendation is created. The different forms of recommendations are explained and a link with recommending music is made. This section ends with trends in music recommendation. The third part of the literature review explains how music is classified and what problems are encountered in doing so. This section is followed with an explanation on how music preferences arise. The next section describes that differences in music preferences could have an effect on the use of the recommendations. The last section makes a link between the recommendations and online advertising. First it is explained how revenues from online advertising are generated, followed by how its effectiveness is measured. The last part of this section describes how online

recommendations influence the effectiveness of online display advertisement.

Motivations for using technology

10 Based on the Theory of Reasoned Action, Davis (1986) created the Technology Acceptance Model (TAM). The TAM model explains what determines the attitude towards new

technology. According to Davis (1986), the factors that influence the attitude are perceived usefulness and perceived ease of use. Ease of use is defined as “the degree to which the prospective users expects the system to be free of effort” (Davis, Bagozzi & Warshaw, 1989, p.985.). So, the ease of use is determined by how efficient a user can perform a task with the specific technology. This need for efficiency has proved to be of major influence on the motivation of people (Bandura, 1982). The efficiency is influenced by external variables such as training, documentation and design of the program for example proper use of menus in the program (Davis, Bagozzi & Warshaw , 1989). Perceived usefulness indicates how well a program helps the user in improving his performance and is also determined by external variables. This is related to ease of use (Davis, Bagozzi & Warshaw, 1989). For example if two systems have an equally strong performance, but one is easier in use, the user will perceive the easiest as more usable. So in order to stimulate people to use technology more intensively, one has to improve its usefulness and make it easy to use.

Online recommendations

When people are exploring and choosing items, they are looking for a balance between search effort and reward. This can lead to situations where the item that has been chosen is

satisfactory, but not optimal (Häubl & Trifts. 2000). When the amount of options from which consumers can choose are relatively small, choice has been proven to be beneficial for the costumer. However, when the amount of option’s given to a consumer is relatively high, consumers will suffer from “choice overload” causing them to get demotivated (Iyengar & Lepper, 2000, p. 1004.). Recommendation systems can help the consumer with decision making that leads to a better reward at reduced effort. This is especially in an online

11 the highest utility, will be recommended to the specific user. In an ideal situation, the user has seen every item and indicated how much (s)he likes a product. However, this is unlikely to happen, leaving gaps in the combination of user preferences and item characteristics.

Therefore, the utility needs to be extrapolated to items with an unknown utility. So the factors that determine the utility need to be found in order to derive a utility for new products. In general, the determination of the utility is done: collaboratively based, content-based or by a hybrid form of collaborative and content-based. (Adomivicius & Tuzhilin. 2005; Åman & Liikkanen, 2017; Häubl & Trifts. 1999; Chung, Rust & Wedel, 2009). Collaborative recommendation systems recommend items that are preferred by users with the same characteristics. So the utility is derived through a comparison with similar users.

Collaborative recommendation systems are often called collaborative filters. Content-based systems recommend items based on similarity between features of items liked previously. The features that content-based music recommenders are often based on are “genre, mood and instrumentation” and can be determined automatically or manually (Åman & Liikkanen, 2017, p. 165.). Hybrid forms between content-based and collaborative approaches combine the two types of recommendations. In the music industry all three types of recommendation systems are used but collaborative filtering is the most popular and present at all the big music streaming platforms (Åman & Liikkanen, 2017).

A common problem with recommendation systems is the cold start problem. This occurs if recommendation system have to recommend products without having data about that person to base the recommendations on. This is likely to happen at the start of using the

recommender system (Dai, Qian, Jiang, Wang & Wu, 2014; Chung, Rust & Wedel, 2009). A problem that collaborative filters suffer from is popularity bias. Because these systems

recommend what is popular with similar users, popular items are recommended and become more popular and unpopular items become even more unpopular (Åman & Liikkanen, 2017).

Recommendation systems are constantly improved. A noticeable trend is the use of contextual factors to determine the utility. One of these contextual factors is group consumption. Hennig-Thurau, Marchand and Marx (2012) found that recommendations for products that are

12 more often searched with terms that describe the need someone wants to satisfy, then with terms that relate to actual features of music.

Kaminskas and Ricci (2012) divide the context for music recommender systems in three categories. The first type is environment related. The environment relates to the time, weather and location that music is listened. An example of an environmental factor is temperature. Temperature has an impact on the emotional state, which can invoke changes in music preference (Kaminskas & Ricci, 2012). The incorporation of environmental factors in music recommendations has become more complex, because the use of mobile devices to listen to music has increased drastically (Kaminskas & Ricci, 2012). This causes an almost infinite amount of situations where music could be listened. The second context type is user related. This type of context refers to the emotional state, activity and demographics of the user. An example of this could be listening music during exercising. The third type of context is multimedia. This refers to the combination of music with other types of media, for example text and images. Such a system would present music that is relevant to the text or image that is looked at.

Music listeners encounter a seemingly unending amount of music nowadays.

Recommendation systems can help consumers with their choice, reducing the effort that it takes to find likable music. This improves the usability of the website. Because usability is one of the main drivers for people to adopt technology, it is assumed that recommendations will lead to an increased use of the platform.

What music are people listing to

13 Music genres can be defined from two different viewpoints: musicological and sociological. The musicological viewpoint categorizes music on the technical properties of a song, such as its structure, melody, or instruments. A factor analysis done by Shäfer and Sedlmeier (2009) indicated that the 25 most common music types can be reduced to 6 dimensions (Table 1).

Style Music types

Sophisticated Jazz, Blues, Swing, Classical Electronic Techno, Trance, House, Dance

Rock Punk, Metal, Rock, Alternative, Gothic, Ska

Rap Hip Hop. Rap, Reggae

Pop Pop, Soul, R'n B, Gospel

Beat, folk, Country

Beat music, Folk, Country , Rock 'n' Roll

Table 1: Music styles

The sociological viewpoint on music genre takes social processes related to a music genre into account, for example the expectation that an artist of a certain music genre shows behaviour that is deemed appropriate for that music genre. The borders for sociological classifications are less strict than for musicological, meaning that different musicological genres can be of the same sociological genre (Lee & Childress, 2016).

For recommendations based on music genre, music is classified based on its musical properties. The link between item characteristics and user characteristics, suggests that content-based recommendation systems are appropriate for this type of recommendations. Collaborative recommendations are based on what similar people like. The focus of these systems is therefore more on the social environment of the users. This type of recommender system would therefore recommend more from a sociological viewpoint.

Music preferences

The conclusion on what is the most popular music genre differs. In the research of Shäfer & Sedlmeier (2009) it is concluded that Rock is the most popular music type followed by Classical and Pop music. North & Hargreaves & Hargreaves (2004) found that Pop music is most listened music type.

14 (1989) conclude that specific technical characteristics of the music, like the tempo and rhythm of a song, are factors that can explain preference for a certain genre. Coming from a more sociological viewpoint, Rentfrow & Gosling (2003) found that personality, self-views and cognitive abilities all correlate with music preference. For example, people who listen to upbeat and conventional music are according to Rentfrow & Gosling (2003) cheerful, socially outgoing, reliable and tend to be relatively conventional. Another predictor for music

preference is age. Age is an important variable because people tend to keep the musical preference they developed during their young adulthood (Holbrook & Schindler, 1989). A recent study from Hemming (2013) replicated the study of Holbrook & Shindler (1989) and found that the age at which the music preference is developed has become younger, going from 23 to 17 years. The personal music preference is also dependent on the environment of a person. This is caused by social influence, which is an important factor in choosing what music to listen to (Salganik, Dodds & Watts, 2006). The majority of people listen to music voluntarily and listen to it in a passive way. For example, listening to music to pass time or as an habit. (North, Hargreaves & Hargreaves, 2004). Shäfer and Sedlmeier (2009) state that music is used to communicate “personal values, ambitions, believes and perceptions” (p. 280.) and that different types of music have different functions. The strength of having a preference for a certain music type is determined by the degree to which this music type fulfils the need that a listener wants to satisfy with listening to this type of music (table 2).

North, Hargreaves & Hargreaves (2004) also conclude that people listen to different types of music for different reasons. Furthermore they emphasise the importance of contextual factors that affect music choice. These contextual factors are: with whom someone is listening to music, where someone is listening music and when the music was listened. For example people prefer other music during a party than during a normal working day. What music people are willing to explore is based on the history and personal preference and is influenced by cultural preconception (North, Hargreaves & Hargreaves, 2004).

Music type Function

Sophisticated Artistic (intellectual) stimulation, identification with artist, no emotional effects, no arousal effects

Electronic Getting energized, getting in an ecstatic mood, understand own feelings Rock Expression of identity

15 Pop Identify with artist, express values

Beat, folk, country

Getting in a good mood, understand own feelings

Table 2: Function per music type

Because people listen to music to fulfil a certain need, it is believed that their previous listening behaviour shows how they fulfil this need. Therefore, it is believed that if the

recommended music is based on previous listening behaviour, the recommender system helps in fulfilling the need of the user. This would make the platform more useful in fulfilling the need of the user and making it more easier to find music of the preferred type. Because ease of use and fulfilling needs are determinants for the use of technology, it is assumed that recommendations based on music preference will improve the attitude towards the publishers website. This would cause an increase in visits and increase the duration of the visits. This leads to the following hypothesis:

H1: recommendations based on music preference are more likely used than generic recommendations.

H2: recommendations based on the music preference will increase the amount of visits to the website.

H3: recommendations based on the music preference will increase the duration of the visits of the website.

Because the recommendations are based on the features of the content (music type), the proposed recommendations are content-based. Content-based recommendations will also prevent popularity bias, since this type o recommendations are not relying on the popularity of songs.

Differences in the use of music recommender system

16 Agreeableness is a personality trait that also is referred to as: friendliness and social

conformity. This trait relates to tolerance, level of cooperation and flexibility of a person. Conscientious people are strong organizer’s and are reliable. Emotional stability is related to how nervous someone is and describes his or her temperament. The fifth personality trait that Barrick and Mount (1991) describe is intellect. This refers to curiosity and how open minded someone is. Intellect is often referred to as: openness to new experiences and is described as the preference for new experiences instead of known routines (Cucu-Ciuhan & Răban-Motounu, 2012; Deng, Liu, Lim & Hu, 2013).

Two of these personality traits could have an effect on the adoption of the recommender systems. The first one is agreeableness. People that score high on this trait are likely to be cooperative. This leads to the assumption that they would be more likely to agree with the music that is recommended to them. The second personality trait of interest is intellect or openness to new experience. This personality trait describes how likely someone is to use something new. Therefore, it is believed that if people score high on this trait, they are less likely to use the recommendations based on music type. The reason for this effect is that the recommended music is of the same type listened before. A user is therefore confronted the same experience as before.

According to the research done by Rentfrow and Gosling (2003), Rap and Electronic music preferences are associated with agreeableness. A preference for Pop music is correlated with openness to new experience. Sophisticated, Rock and Beat, Folk & Country have a high openness to new experience. This leads to the following hypothesis:

H4: People with a preference for Rap, Pop or Electronic music, are more likely to use recommendations based on music type.

Online advertising

17 is therefore often used to increase traffic to a website and to improve sales (Roels &

Fridgeirsottir, 2009). The website that offers space to advertisers are paid when a banner is seen and when it is clicked on (AdSense, 2017). Therefore each user visit can be regarded as an individual product sold by a publisher (Chen & Stallaert, 2014). So an increase in visitors will lead to an increase in revenue. This implies that the effectiveness of the publisher’s website is dependent on its amount of visitors and the willingness of these visitors to click on banners.

Visitors of websites are actively avoiding advertisement. This phenomena is often referred to as banner blindness (Dréze & Hussherr, 2003; Hervet, Guerard, Tremblay & Chtouru, 2011). However, an increase in avoiding banner advertisement does not mean automatically that banner advertisement is not effective. This is because people can memorize the content of advertisement in an explicit and implicit way. Explicit memory refers to the situation where someone is “intentionally and consciously try to recollect a specific past event” (Yang, Roskos-Ewoldsen, Dinu & Arpan, 2006, p. 145.). Most studies on online advertisement effectiveness are based on how well people recall what is advertised and are therefore based on the explicit memory (Hervet, Guerard, Tremblay & Chtouru, 2011). Implicit memory is a process where ‘a specific event can influence the perception and interpretation of subsequent events without recall of the prior event’ (Yang, Roskos-Ewoldsen, Dinu &Arpan, 2006). Research with eye tracking scanners has shown that banner blindness does not mean that people do not see the banner at all, making it still effective through implicit memory, despite that the content is not directly recalled (Hervet, Guerard, Tremblay and Chtouru, 2011).

Click trough rate is the metric that captures how much a banner is clicked on, relative to the number of visitors of the website. Despite the fact that click trough rate is often used as a measurement of effectiveness, it has been criticised for not being accurate. Manchanda, Dubé, Goh & Chintagunta (2006) conclude that if banner advertisement is evaluated on actual purchase behaviour, banner advertisement is effective. The explanation given for low effectiveness of online banner advertisement is that click trough rates do not capture all the effects that is generated by the banner advertisement. According to Ghose &

18 advertisers webpage by clicking on the advertisement.

With click trough rate, only the direct click is measured and active and passive search responses are not determined. Therefore, evaluation by click trough rate makes a banner advertisement look less effective than it really is. Despite that click trough rate is not optimal for determining advertising effectiveness, it is still commonly used as a metric to determine the value of advertisement space (Ghosh & Bhatnagar, 2013). This implies that an increase in click trough rate would increase the value of the advertisement space, making it relevant to investigate if increasing the revenue of advertisement is the goal.

Another interesting aspect that affects effectiveness of banner advertising is duration of the website visit. Ghose and Todri (2015) investigated effects of exposure time on online search behaviour. They found that longer exposure time to advertisement, makes people more likely to direct visit a website, compared to an indirect visit through a search engine. They also report that increasing the exposure time with one minute, increase banner clicks with 0.003.

Elsen, Pieters and Wedel (2016) analysed if exposure duration affects different types of advertisement differently and concluded that exposure duration has an impact on the evaluation of advertisement. They argue that this effect occurs because people need time to identify what brand and which product the advertisement is about and evaluate if it is relevant for them. The findings of their research shows that especially mystery advertisements

(advertisements that does not show what they promote directly) are evaluated more positive if the exposure time is longer. With upfront ads it is immediately clear what brand and product is promoted, these adds are evaluated the same for long and short exposure times. False front advertisements are evaluated more negatively when the exposure time is longer. In false front advertisements the advertised product is associated with a different product of brand. For example by showing the own product in the packaging of a different product. The explanation for this negative effect is that people identified the brand and product, but then find out that the advertisement is about a different brand and product. This forces them to make a schema switch which people do not like. Based on their findings, Elsen, Pieters and Wedel (2016) argue that exposure time is an important feature for optimising advertisement effectiveness.

19 it is believed that longer visits result in a higher click trough rate. This leads to the following hypothesis:

H5: Visit duration has a positive effect on click trough rate.

With the third hypothesis it is proposed that the use of the recommendations will increase the length of visits. Because the sixth hypothesis proposes an effect of duration on click trough rate, it is believed that recommender use will lead to increased click trough rates. This effect is mediated by visit duration, leading to the following hypothesis:

H6: Duration of the website visits mediate the effect that recommendations based on music type have on click trough rate.

Figure 1 is a graphical representations of the conceptual model, with all the hypothesis.

Figure 1: Conceptual model

3. Method

Recommendations based on music type

20 recommends the same songs for every user. These recommendations are located at the top of the website (appendix 1). The recommendations are presented with a picture of the video of the song and the title (appendix 1 & 3). Every week a track is added. There is no rationale behind this song choice, besides that they are relatively new. The recommendations based on music preference will be incorporated in the existing recommendation system to exclude any effects that could be related to a change in presentation of the recommendations.

Every user gets five recommended songs. The users are assigned randomly to either the test group or control group. If a user is assigned to the test group, the user will get

recommendations that are based on his music preference. The recommended songs are from different artist of the same music type. Multiple songs of different artists are chosen because there is a possibility that a user dislikes a certain artist. With different artist the probability that this occurs is smaller.

The songs in the dataset are classified according to the music types of Shäfer and Sedlmeier (2009), resulting into songs classified into Pop, Rock, Rap, Sophisticated, Folk or Electronic. Through a combination of listening to the artist, information from the artists website and the classification used by Spotify, music type is assigned to the artist.

The preference for a specific music type is determined by taking the percentage of how many songs of a specific music type is listened to. The reason for this is to account for differences in the amount of songs that are classified per user. Deriving the preference form counting the numbers of a certain music type would not be reliable since some users might have more songs classified. The users with less songs classified would be classified as having lower preference for a certain music type. By taking the percentage of songs of a certain type that was listened to, this effect is excluded.

21 means that this group will get the same amount of recommendations as the test group, but there are songs of all music types. Because the goal of recommender systems is to help a user to explore items, it is assured that the songs are new to the user. This is done by selecting recently released songs that are not present in the analysed data of the users. The probability that a song was already known by the user and therefore clicked on, is reduced in this way. Appendix 2 is an overview of the recommended songs. The pictures that go with the recommendations can be found in appendix 3.

Finding similar users

Clustering allows to find groups of users that are similar to each other, but are different from other groups of users (Li, NG, Cheung & Huang, 2008). Clustering algorithms are divided into hierarchical clustering and partitional clustering. Hierarchical clustering starts with treating every observation as one single cluster. It then connects the clusters that are closest to each other to form a new cluster. This process is repeated until there is one cluster left, that contains all observation. The interpretation of a hierarchical clustering algorithm is often done with a dendrogram, which is a tree like graph that visualises to which cluster an observation belong (Sharma, Boreocich, Shigemizu, Kamatani, Kubo & Tsunoda, 2017). Partitional clustering methods are designed to find a predetermined number of clusters and the algorithm finds the observations that are closest to the predetermined clusters. An advantage of these methods is that they do not rely on the dendrogram for interpretation, which can be very hard when dataset is large. Disadvantage is that one needs to predetermine the amount of clusters (Zahra, Ghazanfar, Khalid, Azam & Naeem, Prugel-Bennet, 2015).In this research, the six music types of Shäfer and Sedlmeier (2009) serve as the predefined number of clusters. This allows the use of partitional clustering algorithms.

22 A widely used clustering algorithm that suits the needs of this research is the K-means

clustering algorithm designed by MacQueen in 1967. This method is simple, easy to implement and fast. The predefined number of clusters function as centroids in K-means algorithm. With the standard K-means algorithm, the initialization of the centroids is random, leading to inaccurate results (Arthur & Vassilvitskii, 2007). The consequence of this

randomness is that every time the algorithm is used, the results can differ. The K-means++ algorithm proposed by Arthur and Vassilvitskii (2007) is a more accurate and stable version of the original K-means algorithm. With K-means++ the first centroids is also placed randomly. The initialization of the centroids that follow, are based on a probability that is derived from the distance between the centroids. The longer the distance, the higher this probability is (Arthur & Vassilvitskii, 2007). So K-means++ is better than the normal algorithm, because its initializations procedure causes the centroids to be as distinct as possible from each other. Therefore this adaptiation of the standard K-means algorithm is used for the cluster procedure in this research.

The quality of K-means algorithms is determined by the number of clusters it is ordered to find. There are several metrics that support in determining the amount of clusters.

One of the best performing metrics for determining the number of clusters is the Davies and Bourdain (DB) index (Arbalaitz, Gurrutxaga, Muguerza, Pérez & Perona, 2013). The DB index is the result of an evaluation of both intra and inter cluster difference. The lower the index is for a given number of clusters, the better the number of clusters is. When the number is known beforehand, the DB index can be used to validate this (Arbalaitz, Gurrutxaga, Muguerza, Pérez & Perona, 2013). Because the DB index is proven to evaluate the amount of clusters well, it is used to determine the most appropriate amount of music preference clusters.

Recommender adoption

In order to test H1: recommendations based on music type are more effective than the generic recommendations, data on the use of the recommendations is collected. The test group

23 leading to more use of the recommendation system, a logistic binary regression is used, where adoption is the dependent variable. The independent variable is the dummy variable for test group allocation.

Formula: P[Yᵢ = 1] = Λ(α+ T’ᵢ 𝛽1) ᵢ = individual user

Yᵢ = 1 if recommendations are used, 0 if recommendations are not used Tᵢ = 1 if the user is in the test group, 0 if the user is in the control group

Individual differences between clusters

The hypotheses that Pop, Rap and Electronic listeners are more likely to adopt the

recommender system (H4), is tested by means of binary logistic regression. The dependent variable of this regression is adoption. The independent variables is a dummy variable that indicates if the user belongs to the Pop, Rap or Electronic cluster.

Formula: P[Yᵢ = 1] = Λ (α + T’ᵢ 𝛽1+ P’ᵢ 𝛽2 + (T’ᵢP’ᵢ 𝛽3))

ᵢ = individual user

Yᵢ = 1 if recommendations are used, 0 if recommendations are not used Tᵢ = 1 if the user is in the test group, 0 if the user is in the control group

Pᵢ = 1 if the user has a Pop, Rap or Electronic preference, 0 if user has other preference

Increased durations

To investigate if the recommendations have led to an increase in the duration of a visit (H3), a transformation from time to session length is made. The dataset does not provide how long a visit of a user was. This is derived by adding up the duration of the played songs for every session. The length of a song is calculated by subtracting the time that a song is played, with the time that a previous song is played at. For songs that are at the end of a user’s session and therefore have no song the can be subtracted from, the duration is added manually, by looking up the song on YouTube. The total sessions are divided by the amount of sessions in order to get to an average session length per user. The third hypothesis is tested by means of linear regression. The calculated average duration per users is used as the dependent variable. The independent variable is a dummy variable for test group allocation.

24

ᵢ = individual user

Yᵢ = average duration of a user

Tᵢ = 1 if the user is in the test group, 0 if the user is in the control group 𝜀ᵢ = error term

Effects of visit duration and recommender use on click trough rate

In order to test if there is a mediating effect of duration on click trough rate (H6), the non-parametric bootstrap mediation analysis proposed by Preacher and Hayes (2004) was used. This method is chosen because a small sample (< 25), is likely to violate the assumption of a normal distribution (Preacher & Hayes, 2004). In this analysis the dependent variable is click trough rate. The click trough rate of every day of the test period is calculated. The click trough rate is the result of the amount of times that banners are clicked on, divided by the amount of sessions. The mediator is the average duration of a session for every day in the test period. This is calculated by summing the session length and dividing it by the amount of session of that specific day. This mediation analysis also shows the direct effect of visit duration on click trough rate and is therefore also used to test H7: Visit duration has a positive effect on click trough rate.

Formula: yᵢ = α + 𝛽5Tᵢ + 𝛽6Dᵢ + 𝛽7(TᵢDᵢ) + 𝜀ᵢ

ᵢ = individual user Yᵢ = click trough rate

Tᵢ = 1 if the user is in the test group, 0 if the user is in the control group Dᵢ = visit duration

𝜀ᵢ = error term

4. Results

Classification

25 and only 1,3% smaller than the Rap. Therefore Dutch Folk music is regarded as important and taken as an extra music type.

Music type Amount of songs % % cumulative

Pop 6959 46,8 46,8 Rock 2918 19,6 66,4 Electronic 1720 11,6 77,9 Rap 1295 8,7 86,6 Dutch Folk 1108 7,4 94,1 Folk 448 3,0 97,1 Sophisticated 436 2,9 100,0 Total 14884 100,0

Table 3: Number of classifications per music type

Clustering

The K-means++ cluster analysis has resulted in 6 distinct clusters (table 4) Cluster name Cluster number % Pop % Rock % Electronic % Sophisticated % Dutch % Folk % Rap Size Electronic 1 33 8 47 1 3 1 8 33 Pop 2 72 8 9 2 1 1 7 84 Rock 3 23 58 4 4 3 5 3 43 Rap 4 31 8 15 1 5 0 40 25 Dutch Folk 5 24 10 4 2 52 4 3 19 Pop/Rock 6 45 20 8 7 8 6 5 97

Table 4: Kmeans++ cluster results

26 Dutch Folk music resulted in a distinct cluster with a strong preference for Dutch Folk music. This cluster also has a less strong preference for Pop music. Cluster six prefers Pop music and Rock music. Given that there is for both Pop and Rock already a cluster that singly prefers Pop or Rock (cluster 2 and 3), it is assumed that the people in this cluster are really looking for the combination of these two types. Therefore the sixed cluster is regarded as a

combination of Pop and Rock. A preference for Sophisticated music or Folk music could not be observed from the cluster results. When comparing the size of the different clusters, it is observed that the two biggest clusters are the Pop/Rock and Pop cluster. Together these clusters are accounting for 60%, of the total dataset. The third largest cluster is Rock music. This type of music is preferred by 14% of the users. When adding these sizes, it is observed that a majority of 74% of the users listens to Pop, Rock or a combination of these two types.

The lowest DB score was generated at six clusters, with a DB index of 33. Five clusters result in a DB index of 76 and seven clusters generate an index of 42. After seven clusters the DB indexes expand rapidly. Eight clusters generate a DB index of 109 and at nine clusters the DB index is already at 137.

Summary of the collected data

Over a period of 7 days the effect of recommendations on 301 users was measured. From 189 of the users, information could be analysed, since not every user has been using the platform in the seven day testing period. The response per cluster is summarized in table 5. The recommendations are used 32 times in total. Based on this response data H1, H2, H3 and H4 are tested.

Cluster Size % Total

sessions Average Amount of sessions Average session length (seconds) Number of recommender uses Banner clicks Electronic 14 7,4 99 7.1 2331 7 0 Pop 45 23,9 418 9.3 4180 22 2 Rock 19 10,1 192 10.1 3382 0 0 Rap 14 7,4 81 5.7 4575 0 0 Dutch 13 6,9 81 6.2 4198 0 1 Pop/Rock 44 23,4 341 7.8 3847 3 2 Control 39 20,7 229 5.8 3991 0 3 Total 188 100 1441 7.4 3786 32 8

27

Effects of music type based recommendations

The logistic binary regression to test if recommendations based on music type are more likely to be used (H1), have not led to significant results (P-value = 0.992, std. error = 1722.02, β = 17.04).

Amount of visits

The linear regression used to test if recommendations based on music type are increasing the amount of visits (H2), did not show a significant effect (P-value = 0.111, std. error = 1.413, β = 2.26, R² = .0004).

Effect of different music preferences

Although that there is no significant effect that proves that recommendations based on music type are more likely to be used then generic recommendations, there are differences between music preferences (H4). From the summary of the dataset (table 6), one can already observe that the Pop and Electronic listeners have used the recommendations more often. The logistic binary regression showed that the group: Pop, Electronic and Rap, are more likely to use the music type recommendations (β = 1.649, std. error = .800, p = .039). From the exponent of the estimate, it is derived that the a user is 5.2 times more likely to use the music type recommendations if s(he) belongs to this group. However, the Rap cluster did not use the recommendations and the Electronic cluster used it three times less than the Pop cluster. A logistic binary regression is performed to see if the strength of the effect per preference differs. The effect of the Electronic preference was not significant (β = 0.232, std. error = 1.087, p = .831). The Pop preference showed a highly significant effect (β = 1.985, std. error = .703, p = .004). The likelihood of using the recommendations based on music type is 7.3 times higher when someone has a Pop preference. This shows that the significant effect of having a preference for Electronic, Rap or Pop music, is caused by the strong effect of Pop music preference.

Increase in click trough rates

28

5. Discussion

This research was designed to determine if recommendation systems can increase the revenue of advertisement based websites, through higher click trough rates and an increased amount of visitors. An increase in click trough rate is expected because recommendations increase the ease of use and usability of the music platform; making visitors use the platform more often and for longer periods. More often use would lead to more banner impressions. Longer visits of the website increase the probability of banner clicks (Ghosh & Todri-Adampoulos, 2015). In order to test these hypotheses, content-based recommendations based on the music types of Shäfer and Sedlmeier (2009), where created. From 188 users, the use of the recommendations and the impact on click trough rate was measured. Of the 188 users, 33 users received generic recommendations that are not based on their music preference. This group functions as a control group. The rest of the users received recommendations based on their preferred music type. This preference was derived from their listening behaviour. The collected data has led to the following findings and conclusions.

The results show that the probability of using recommendation system based on music type is not significantly different from the probability of using generic recommendations. So there is no support for H1. The data did not show significant effects that support H2:

29 good mood, which in case would lead to being less thoughtful about music selection and therefore less critical with music selection. This could explain why there is an increased probability to use the recommendations for Pop listeners.

The findings of this research, lead to the conclusion that recommendations based on music type do not increase the amount of visits and do not increase the clicks on banners. Therefore, content-based recommendations do not lead to an increase of advertisement revenues. Despite that no support for the main hypothesis was found, there are some additional findings to report.

The most listened music types are Pop, Rock and a combination of Pop and Rock music. This is in line with research of Shäfer and Sedlemeier (2009), where Rock music was the most popular. It is also in line with the research of North, Hargreaves and Hargreaves (2004), where Pop music was the most popular music type. No distinct clusters could be found for the Sophisticated music type as well as for the Folk music type. What is distinctly different from previous research is the presence of Dutch Folk music. 7,4 percent of the songs were

classified as Dutch Folk, making this more popular than Sophisticated music with 3 percent and folk music with 2.9 percent. The K-means++ clustering algorithm led to a distinct cluster for Dutch Folk listeners. In this cluster 52 percent of the songs was Dutch Folk music

30 Because the amount of sessions of the clusters that used the recommendations are higher than the non-users (table 5), the effect of session amount on the use of recommendations was investigated. With a logistic binary regression, a marginally significant effect is found. (β = .050, std. error = .264, p = .054). The odds of using the recommendation system increase with five percent for every session that is added. A possible explanation for this observation might come from the research of Goldberg, Chattopadhyay, Gorn and Rosenblatt (1993), were it is concluded that repeated exposure can result in boredom. If a user is bored with the music (s)he is listening, (s)he might start exploring new music and become more likely to us the recommendations.

The results of this research have led to the conclusion that the implementation of recommendation system, to recommend music based on music type, does not affect the parameters on which the revenue of free to use music streaming platforms depend. This conclusion helps in the design of music streaming platforms by discouraging a system as proposed in this research, if revenue maximisation through advertising is the main goal. Since Pop listeners turned out to be the only cluster that where likely to use the recommendations, only recommending Pop music might be as efficient as adapting the recommendations to six different music types. If the recommendations are selected manually as with this research, the work effort would decrease to a sixed of creating recommendations for all the preferences. The additional result showed that Electronic user tend to listen in shorter sessions and that this might be caused by the context in which this music is listened. This result helps in the

creation of recommendations based on context. In this case this context would be working out. Both Åman & Liikkanen (2017) and Kaminskas & Ricci (2012) emphasise the importance of context with music recommendations. More sessions increased the odds of using recommendations. It is believed that this might be caused by an increase in boredom with the current music, making the user more likely to start looking for new music. This feature could be used for different treatment of the listeners with increased session. For example recommending more music to them.

31

6. Limitations and future research

Because this research is performed in a short period of time, the collected data is limited in its quantity. The results of this research might differ if the recommendations would have been online for a longer period then seven days and with more sets of recommendations. Therefore the generalizability of this research is questionable. Future research should test the proposed method for a longer period of time and with more recommendations. This would strengthen the conclusion of this research.

The type of advertisement that the user encountered could not be controlled, because the platform used Google Adsense. This is a third party that links advertisers to publishers based on the web history of the visitor (AdSense, 2017). So every visitor gets different

advertisements. The impact of duration on advertisement can differ per type of advertisement (Elsen, Pieters & Wedel, 2016).

Because a large amount of songs where Dutch Folk songs and one cluster had a distinct preference for this type of music, Dutch Folk music was regarded as a separate music type. No literature was found about this music type. Therefore there is no support that this music type is distinctly different from the other music types. The preference for Dutch Folk music could therefore not be related to the fulfilment of a certain need, which is the argument for choosing music type to base the recommendations on.

Only the most popular music was classified. This could have made it possible that users are classified differently when every song is classified. For example, if a user listens to popular Pop music but also to niche Rock music, all the Pop songs would be classified, but the Rock songs not. The user is then classified as a Pop listener, instead of a Pop/Rock listener. Music was classified according to the genre of music an artist produces. However, artists experiment with musical genres. So there might have been songs that are classified to a certain genre, but are of a different genre because artist are not bounded by genre.

32 There are several types of recommendation systems. In this research only content-based recommendations are used. Future research should also investigate the effects of different types of recommender systems, such as collaborative recommendations or hybrid

recommendations. This is important because the bases on which the recommendations are done, differ dramatically between the different type of recommendation systems. For example in a collaborative filter social features would be incorporated, because the popular items of similar users are recommended.

33

6. Reference list

Adomavicius, G., & Tuzhilin, A. (2005). Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions. IEEE transactions on

knowledge and data engineering, 17(6), 734-749.

Ajzen, I., & Fishbein, M. (1975). Belief, attitude, intention and behavior: An introduction to theory and research.

Arbelaitz, O., Gurrutxaga, I., Muguerza, J., Pérez, J. M., & Perona, I. (2013). An extensive comparative study of cluster validity indices. Pattern Recognition, 46(1), 243-256.

Arthur, D., & Vassilvitskii, S. (2007, January). k-means++: The advantages of careful seeding. In Proceedings of the eighteenth annual ACM-SIAM symposium on Discrete

algorithms (pp. 1027-1035). Society for Industrial and Applied Mathematics.

Anderson, C. 2009. FREE: The future of a Radical Price. New York: Hyperion Books. Bandura, A. (1982). Self-efficacy mechanism in human agency. American psychologist,

37(2), 122.

Chen, J., & Stallaert, J. (2014). An economic analysis of online advertising using behavioral targeting. Mis Quarterly, 38(2), 429-449.

Cho, C. H. (2004). Why do people avoid advertising on the internet?. Journal of advertising,

33(4), 89-97.

Chung, T. S., Rust, R. T., & Wedel, M. (2009). My mobile music: An adaptive personalization system for digital audio players. Marketing Science, 28(1), 52-68.

Cucu-Ciuhan, G., & Răban-Motounu, N. (2012). The Openness to Experience Questionnaire: construction and validation. Procedia-Social and Behavioral Sciences, 33, 717-721.

34 Davis, F. D. (1986). A technology acceptance model for empirically testing new end-user

information systems: Theory and results (Doctoral dissertation, Massachusetts Institute of

Technology).

Davis, F. D., Bagozzi, R. P., & Warshaw, P. R. (1989). User acceptance of computer

technology: a comparison of two theoretical models. Management science, 35(8), 982-1003.

Deng, S., Liu, Y., Li, H., & Hu, F. (2013). How does personality matter? An investigation of the impact of extraversion on individuals' SNS use. Cyberpsychology, Behavior, and Social

Networking, 16(8), 575-581.

Drèze, X., & Hussherr, F. X. (2003). Internet advertising: Is anybody watching?. Journal of

interactive marketing, 17(4), 8-23.

Elsen, M., Pieters, R., & Wedel, M. (2016). Thin Slice Impressions: How Advertising Evaluation Depends on Exposure Duration. Journal of Marketing Research, 53(4), 563-579.

Fan, M., Kumar, S., & Whinston, A. B. (2007). Selling or advertising: strategies for providing digital media online. Journal of Management Information Systems, 24(3), 143-166.

Finn, R., & Wadhwa, K. (2014). The ethics of “smart” advertising and regulatory initiatives in the consumer intelligence industry. info, 16(3), 22-39.

Finnäs, L. (1989). How can musical preferences be modified? A research review. Bulletin of

the Council for Research in Music Education, 1-58.

Fleder, D., & Hosanagar, K. (2009). Blockbuster culture's next rise or fall: The impact of recommender systems on sales diversity. Management science, 55(5), 697-712.

35 Ghose, A., & Todri, V. (2015). Towards a Digital Attribution Model: Measuring the Impact of Display Advertising on Online Consumer Behavior.

Ghosh, H., & Bhatnagar, A. (2013). On measuring and increasing the effectiveness of banner advertising. MIS Review, 19(1), 25-44.

Goldberg, L. R. (1993). The structure of phenotypic personality traits. American

psychologist, 48(1), 26.

Goldberg, M. E., Chattopadhyay, A., Gorn, G. J., & Rosenblatt, J. (1993). Music, music videos, and wear out. Psychology & Marketing (1986-1998), 10(1), 1.

Goldstein, D. G., Suri, S., McAfee, R. P., Ekstrand-Abueg, M., & Diaz, F. (2014). The economic and cognitive costs of annoying display advertisements. Journal of Marketing

Research, 51(6), 742-752.

Häubl, G., & Murray, K. B. (2001). Recommending or persuading?: the impact of a shopping agent's algorithm on user behavior. In Proceedings of the 3rd ACM conference on Electronic

Commerce (pp. 163-170). ACM.

Häubl, G., & Trifts, V. (2000). Consumer decision making in online shopping environments: The effects of interactive decision aids. Marketing science, 19(1), 4-21.

Hemming, J. (2013). Is there a peak in popular music preference at a certain song-specific age? A replication of Holbrook & Schindler’s 1989 study. Musicae Scientiae, 17(3), 293-304.

Hennig-Thurau, T., Marchand, A., & Marx, P. (2012). Can automated group recommender systems help consumers make better choices?. Journal of Marketing, 76(5), 89-109.

Hervet, G., Guérard, K., Tremblay, S., & Chtourou, M. S. (2011). Is banner blindness

genuine? Eye tracking internet text advertising. Applied cognitive psychology, 25(5), 708-716.

36 IFPI (2017). Retrieved on: 8-6-2017. URL: http://www.ifpi.org/facts-and-stats.php

Iyengar, S. S., & Lepper, M. R. (2000). When choice is demotivating: Can one desire too much of a good thing?. Journal of personality and social psychology, 79(6), 995.

Kaminskas, M., & Ricci, F. (2012). Contextual music information retrieval and

recommendation: State of the art and challenges. Computer Science Review, 6(2), 89-119.

Kim, J. Y., & Belkin, N. J. (2002, October). Categories of Music Description and Search Terms and Phrases Used by Non-Music Experts. In ISMIR (Vol. 2, pp. 209-214).

Knobloch, S., & Zillmann, D. (2002). Mood management via the digital jukebox. Journal of

communication, 52(2), 351-366.

Manchanda, P., Dubé, J. P., Goh, K. Y., & Chintagunta, P. K. (2006). The effect of banner advertising on internet purchasing. Journal of Marketing Research, 43(1), 98-108.

North, A. C., Hargreaves, D. J., & Hargreaves, J. J. (2004). Uses of music in everyday life.

Music Perception: An Interdisciplinary Journal, 22(1), 41-77.

Preacher, K. J., & Hayes, A. F. (2004). SPSS and SAS procedures for estimating indirect effects in simple mediation models. Behavior research methods, 36(4), 717-731.

Rentfrow, P. J., & Gosling, S. D. (2003). The do re mi's of everyday life: the structure and personality correlates of music preferences. Journal of personality and social psychology,

84(6), 1236.

Salganik, M. J., Dodds, P. S., & Watts, D. J. (2006). Experimental study of inequality and unpredictability in an artificial cultural market. science, 311(5762), 854-856.

Schäfer, T., & Sedlmeier, P. (2009). From the functions of music to music preference.

37 Shao, G. (2009). Understanding the appeal of user-generated media: a uses and gratification perspective. Internet Research, 19(1), 7-25.

Sharma, A., Boroevich, K. A., Shigemizu, D., Kamatani, Y., Kubo, M., & Tsunoda, T. (2017). Hierarchical maximum likelihood clustering approach. IEEE Transactions on

Biomedical Engineering, 64(1), 112-122.

Silver, D., Lee, M., & Childress, C. C. (2016). Genre Complexes in Popular Music. PloS one,

11(5).

Smith, A & Anderson, M. 2016. Online Shopping and E-Commerce. PEW Research Center, URL: http://www.pewinternet.org/2016/12/19/online-reviews/

Statista (2017). Statistics and facts about YouTube. Retrieved on: 10-04-2017. URL: https://www.statista.com/topics/2019/youtube/

Yang, M., Roskos-Ewoldsen, D. R., Dinu, L., & Arpan, L. M. (2006). The effectiveness of" in-game" advertising: Comparing college students' explicit and implicit memory for brand names. Journal of Advertising, 35(4), 143-152.

Zahra, S., Ghazanfar, M. A., Khalid, A., Azam, M. A., Naeem, U., & Prugel-Bennett, A. (2015). Novel centroid selection approaches for KMeans-clustering based recommender systems. Information Sciences, 320, 156-189.

Åman, P., & Liikkanen, L. A. (2017). Interacting with Context Factors in Music

Recommendation and Discovery. International Journal of Human–Computer Interaction,

38

7. Appendix

39

Appendix 2: Overview of the recommended songs per cluster:

Cluster number Cluster name Song type Song title youtube id / picture name Duration/ sec

1 electronic electronic Kygo ft. Selena Gomez - It Ain't Me u3VTKvdAuIY 242 1 electronic electronic Martin Garrix ft. Brooks - Byte zH9sXggRlxc 231 1 electronic electronic Lady Bee ft Jalise Romy - Rebel BY8q7n7-nH0 176 1 electronic pop Katy Perry ft. Skip Marley - Chained To The Rhythm Um7pMggPnug 241

1 electronic pop Rochelle - You Got Something SgLbO2bWQD8 187

2 pop pop Katy Perry ft. Skip Marley - Chained To The Rhythm Um7pMggPnug 241

2 pop pop Rochelle - You Got Something SgLbO2bWQD8 187

2 pop pop Louisa Johnson - Best Behaviour 1sA7Ojgzzus 215

2 pop pop Ed Sheeran - Galway Girl 87gWaABqGYs 200

2 pop pop Charlie Puth - Attention nfs8NYg7yQM 232

3 rock rock Royal Blood - Lights Out ZSznpyG9CHY 242

3 rock rock Metallica - Hardwired uhBHL3v4d3I 199

3 rock rock Volbeat - Seal The Deal _v2J_stvHes 272

3 rock rock Rise Against - House On Fire Ui1eRqX0mZo 195

3 rock pop Katy Perry ft. Skip Marley - Chained To The Rhythm Um7pMggPnug 241 4 rap rap 2 Chainz ft. Ty Dolla $ign, Trey Songz, Jhen‚ Aiko –

It's A Vibe

tU3p6mz-uxU 200

4 rap rap Kendrick Lamar - DNA NLZRYQMLDW4 286

4 rap rap SBMG - Laag/Boven ft. Latifah (prod. MAFQEES) FA9TRc2KOEw 212 4 rap pop Katy Perry ft. Skip Marley - Chained To The Rhythm Um7pMggPnug 187

4 rap pop Rochelle - You Got Something SgLbO2bWQD8 176

5 dutch dutch Klubbb 3 - Het leven danst sirtaki ZX9jcHPKba0 214

5 dutch dutch Tim Douwsma - Dat Ding Met Jou MEWavNhjWY4 211

5 dutch dutch Dries Roelvink - Hermanos ZO2M7WyGIgk 172

5 dutch pop Katy Perry ft. Skip Marley - Chained To The Rhythm Um7pMggPnug 241

5 dutch pop Rochelle - You Got Something SgLbO2bWQD8 242

6 pop/rock pop Katy Perry ft. Skip Marley - Chained To The Rhythm Um7pMggPnug 241

6 pop/rock pop Rochelle - You Got Something SgLbO2bWQD8 187

6 pop/rock pop Louisa Johnson - Best Behaviour 1sA7Ojgzzus 242

6 pop/rock rock Metallica - Hardwired uhBHL3v4d3I 199

6 pop/rock rock Royal Blood - Lights Out ZSznpyG9CHY 242

7 control pop Katy Perry ft. Skip Marley - Chained To The Rhythm Um7pMggPnug 241

7 control rock Royal Blood - Lights Out ZSznpyG9CHY 242

7 control electronic Lady Bee ft Jalise Romy - Rebel BY8q7n7-nH0 176

7 control ned Tim Douwsma - Dat Ding Met Jou MEWavNhjWY4 211

7 control rap Kendrick Lamar - DNA NLZRYQMLDW4 286

40