Temporal Dynamics in Music Listening Behavior: A Case Study of Online Music Service

9th IEEE/ACIS International Conference on Computer and Information Science Temporal Dynamics in Music Listening Behavior: A Case Study of Online Music Service Chan Ho Park Division of Technology and Development Neowiz Bugs Corporation Seoul, Korea chpark@neowizbugs.com Minsuk Kahng School of Computer Science and Engineering Seoul National University Seoul, Korea minsuk@europa.snu.ac.kr Abstract Although temporal context may significantly affect the popularity of items and user preference over items, traditional information filtering techniques such as recommender systems have not sufficiently considered temporal factors. Modeling temporal dynamics in user behavior is not trivial, and it is challenging to study its effect in order to provide better recommendation results to users. To incorporate temporal effects into information filtering systems, we analyze a large sized real-world usage log data gathered from Bugs Music, which is one of the well-known online music service in Korea, and study temporal dynamics in users music listening behaviors considering periodicity of time dimension and popularity change. We insist that the result of our analysis can be a useful guideline to the industry which delivers music items to users and tries to consider temporal context in their recommendations. Keywords-temporal dynamics; data analysis; usage log analysis; popularity; recommendation; context-awareness; online music; I. INTRODUCTION As the amount of information available on the Web increases tremendously, it has become more difficult for users to choose appropriate contents or items. To help users tackle this information overload, a wide range of recommender systems have been proposed and successfully applied to various Web based services (e.g. Amazon 1, Netflix 2, Last.fm 3, etc) that deliver contents or items to users who may like them. The simple but powerful way of generating a recommended item list is to show popular items to all users. For example, the Billboard chart can be an example of a recommendation list that is provided weekly, based on popularity of the songs. However, Wu et al. argued that a simple popular item list is not a good strategy to increase the click-through rate [1]. They concluded that service provider should consider both popularity and novelty to maximize users attention. Although there have been few approaches to incorporating temporal dynamics to their systems, most of the traditional 1 Amazon, http://www.amazon.com 2 NetFlix, http://www.netflix.com 3 Last.fm, http://last.fm recommender systems do not consider time as an important factor; however, in a practical situation, there exists temporal dynamics, also known as concept drift [2], [3]. To consider temporal factors for incorporating them to information filtering systems, we need to consider following issues. First of all, we need to consider that new items are continuously appearing, and their popularity also changes. Thus, systems should detect new items and consider their popularity changes. Moreover, user preference over items is also changing [3], [4]. For instance, a user interested in Pop Rock music might prefer Hip-hop style a year later. In addition, Koren showed that users overall rating scale also can change, by analyzing Netflix dataset [2]. Among many contextual dimensions, such as time, location, user, temperature, etc., the time dimension is distinguishable in that developing a model since timestamp is not just simple continuous or categorical variables. Timestamp is continuous and periodic at the same time. Therefore, preference on some contents or items can be also periodic [5], [6]. For example, Christmas carols are listened more often in December and queries related to movie are requested more on weekends. There are various approaches to consider these kinds of periodic effects [6] [8]. In this paper, we analyze the usage log data achieved from Bugs Music service especially focusing on temporal dynamics in the logs. We examine how temporal context affects users preference on songs and the popularity of songs changes since they are released. We insist that the result of our analysis can give useful insight to other researchers and the industry that aim to recommend music items considering temporal context. The rest of the paper is organized as follows. We present related work in Section 2, and Section 3 describes our dataset and how we sampled and cleaned the data. In Section 4, we analyzed our dataset considering several aspects. First, we present basic data analysis and show the periodic characteristics. Next, we show how temporal context affects users preference on music. Also, we try to figure out how the popularity of music changes over time. We conclude the paper in Section 5. 978--7695-4147-1/1 $26. 21 IEEE DOI 1.119/ICIS.21.142 573

II. RELATED WORK In this section, we review the literature about modeling temporal effects for delivering contents to users. Collaborative filtering is one of the most successful techniques for recommending items to users [9]. However, it gives the same weight to each log record whether it is from one year or one day ago. Many researchers tried to give them different effect with respect to recency [1], [11]. In addition, traditional methods assume users preferences are static and not changing. There are a number of approaches to capturing the drift of users interests [3], [4]. Koren incorporated time factor into collaborative filtering model [2]. He assumes that user s rating scale, item popularity, and users preference for items can change over time. This time changing collaborative filtering model describes Netflix dataset well. There are also several approaches to capturing periodic effects. Vlachos et al., represented time series as a linear combination of complex sinusoids to find periodic queries in Web search [6]. Alonso et al. used four granules which are days, weeks, months, and years to represent time [8]. Park et al. [12] used a fuzzy membership function to categorize continuous dimension for context-aware music recommendation using Bayesian Network, and Shin et al. [7] applied the idea of using the fuzzy membership function to represent timestamp as contextual concepts for context abstraction (especially for timestamp). A. Data Source III. DATA PREPARATION The usage log data we used in this paper is from Bugs Music 4. It is one of the largest commercial online music services in Korea. The data is collected from August 28 to July 29, a period of almost one year. The main data source of our analysis is only the usage log table. Although we can look up the metadata table to figure out the meaning of results, the metadata is not our main data source. Each log record in the log table contains user identification, song identification, timestamp, length of service in seconds, etc. In other words, one record corresponds to one event which means a user listened to a song at a particular time. Table I shows an example of the log table. Table I EXAMPLE OF USAGE LOG TABLE No. Timestamp User ID Song ID Svc. Len. (s) 1 29-7-1 18:3: 11 23456 242 2 29-7-1 18:3:3 651 15348 15 3 29-7-1 18:3:5 3891 57354 211 - - - - - 9999 29-7-1 18:1:22 651 57354 211 4 Bugs Music, http://www.bugs.co.kr B. Sampling Strategy The original log table contains billions of records. Since our aim is to get an approximate distribution of popularity, we sampled the dataset. First, we sampled users. Based on user identification number, we extracted about 1 percent of users. In addition, we sampled 1 percent of records based on timestamp value. In other words, the sampled data is about 1 percent of the whole log dataset. Actually, the data we used are less than 1 percent since we went through a data cleansing process which we will explain in the next subsection. Table II shows basic statistical information after data sampling and cleansing. Table II BASIC COUNT INFORMATION Number of Log Records 14,85,737 Number of Tracks (Songs) 322,85 Number of Users 71,112 Number of Dates 363 Since our analysis does not include join operation between data records, this kind of sampling strategy does not affect the result of our analysis. C. Interpreting Implicit Behavior for Data Cleansing Since raw log records have noisy data, we need to remove some records which have inaccurate or inappropriate information. The log data which does not reflect users actual preference needs to be removed. In the music context, observing skipping behavior is a great way of determining user preference [13]. The log record of a user listening to a very small portion of the whole track, in other words, skipping, shows the user did not want to listen to music in her context. We removed the log records in which the length listened is less than half of whole length. Moreover, the records whose service length is much (2 percent) more than track length are removed since it may show users stopped listening to that music in the middle of the track, or some error caused by external system issues. A. Basic Analysis IV. ANALYSIS Before we examined the effect of temporal dynamics, we performed some basic data analysis. 1) Users: Figure 1 displays the number of music listening events for users who have more than 1 listening events. The horizontal axis represents users and aligned by the number of events. The shape of distribution looks similar to a typical Web usage distribution [14]. Figure 2 shows the relationship between the number of music listening events for users and number of unique songs listened for users. Generally, the contents, such as 574

news, blog, video, are consumed just once for each user. Thus, service provider can use binary preference for these domains [15]. However, in the music domain, people listen to songs repeatedly. If a user likes particular songs, she adds them to her playlist, and listens to them again and again. Thus, the service provider should carefully interpret these kind of repeated events under usage log. The dots at the right side of the plot in Figure 2 means people listened to particular songs many times, or they do not tend to change their playlist. is a tendency that the number of listening events per user for teenage idols songs is relatively high. Therefore, when service providers give ranking of popular music, simple counting cannot reflect all users preferences. To overcome this problem, they can use the number of distinct users who listened to that song. Otherwise, they should apply some other techniques to handle this bias. 1 # of listeners 1 1 1 or listeners 1 1 1 1 1 1 1 1 1 1 song 1 1 1 1 1 1 user Figure 1. Number of Music Listening Events for Users 1 Figure 3. B. Periodicity Number of Listening Events and Listeners for Songs In this subsection, we present some time-series analyses, and show several kinds of periodic patterns in our dataset. Figure 4 presents the number of music listening events for each day over one year. We can see periodic patterns the lines are going up and down every 7 days. This is because user traffic is decreased on the weekend. # of listeners 1 6 5 in a day moving average 1 1 1 1 1 4 3 2 Figure 2. for Users Relationship between Number of Listening Events and Songs 2) Songs: Figure 3 shows a basic analysis for songs. The upper one (plus sign in red) indicates the number of listening events for songs and the lower one (x sign in green) indicates the number of distinct users who listened to that song. We can see that people listened to popular songs about 5 times on average. The value decreases as its popularity decreases. In addition, there are some songs that people listened to more than 5 times on average. These songs are plotted below in the figure. We found that there 1 Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug date Figure 4. Number of Listening Events for each day Figure 5 represents the number of events using box-andwhisker plot for each day of the week, which shows that traffic decreases significantly on the weekend. In addition to weekly periodic patterns, there is also a daily cycle. We plot the number of listening events for each 575

hour in Figure 6. The distribution shows that traffic reaches the peak around 5 PM, and the number of records at 5 AM is about 15 percent of that at 5 PM. In fact, the users in global online services, such as Yahoo, are from various time zones [5]. Although the users in our dataset are not from the same time zone, most users are located in a particular time zone (GMT+9:). per day per day 52 5 48 46 44 42 4 38 36 34 32 3 Figure 5. 35 3 25 2 15 1 5 box plot avg Mon Tue Wed Thu Fri Sat Sun day of week Number of Listening Events for each day of week box plot avg 1 2 3 4 5 6 7 8 9 1 11 12 13 14 15 16 17 18 19 2 21 22 23 hour Figure 6. C. Temporal Context Number of Listening Events for each hour Nowadays, there is a lot of effort to incorporate contextual information into applications, such as search engine results, recommender systems [16], [17]. In the music domain, we have some stereotype related to temporal context, such as people prefer dance music more in the summer, people tend to listen to sad music at night than in the morning. Shin et al. showed that the quality of recommendation can be improved by considering temporal context, such as summer, night, weekend, etc [7]. We used genre to classify songs using our 2-level genre metadata table. Although Shin et al. mapped songs to concepts using tags and external sources, such as ODP 5, we do not have this kind of information, so we use genre in this paper. First, we want to figure out whether time of day affects users music preference. We picked 1 genres to show the difference between them. We normalized the values to recognize the differences more easily. Figure 7 shows the result. rate of listening events.12.1.8.6.4.2 Ballad Rock Ballad Dance Pop Pop Rock RNB Adult contemporary Indie Rock Hip-Hop Solo Instrumental Kids 1 2 3 4 5 6 7 8 9 1 11 12 13 14 15 16 17 18 19 2 21 22 23 hour Figure 7. Daily Effect in Genre Preference Some minor genres including kids and solo instrumental show different patterns compared to most genres, such as ballad, dance pop, and rock. The traffic of these genres has a peak at noon, and decreases rapidly in the evening, while most genres have higher values in the evening. Among popular genres, there is also a slightly different pattern. For instance, people prefer listening to rhythm-and-blues to dance pop at night. Next, we try to see if there exists seasonal effects on music preference. Figure 8 expresses changes in genre preference 5 ODP - Open Directory Project, http://www.dmoz.org rate of listening events.4.35.3.25.2.15.1.5 Ballad Dance Pop RNB Hip-hop TV Drama Pop-Rap Trot Pop-Rock Carol 1 2 3 4 5 6 7 8 9 1 11 12 month Figure 8. Seasonal Effect in Genre Preference 576

throughout a year. We note several findings in this figure. Ballad songs are preferred over dance pop, and the difference is relatively big in the winter. On the other hand, people like to listen to dance-pop music more in the summer (June and July). One of the interesting results is Christmas carol. While there is a very small number of listening events (less than.5 percent) from January to November, 4 percent of listening events is related to Christmas carol in December. D. Popularity Change In this subsection, we analyze how popularity of songs changes since they are released. To simplify our analysis, we pick some popular songs. Figure 9 shows their daily popularity (number of listening events for each day) and each line represents one song. To compare the pattern for change between songs, we let the horizontal axis represent the number of days after the song is released, or age of songs. As we can see Figure 1, the patterns for songs are similar. The distribution may fit into several diffusion models. The diffusion models, such as Bass diffusion model [18], describe how new products get adopted and can be used to predict the popularity of products. Compared to other kinds of offline products, the characteristics for online songs reach the peak very fast. Therefore, it is important to detect songs in which the rate of increase in popularity is high. To push them to users, Bugs already provides a list of real-time song ranking service. This kind of model also makes it possible to predict the cumulative number of listening events from the early listening history of the song [19]. 12 1 8 6 4 2 Track #1 (Nobody) Track #2 (As I was shot) Track #3 (Day by Day) Track #4 (Gee) Track #5 (Mirotic) Track #6 (Can you hear me?) Track #7 (Love... him) Track #8 (Rain and You) Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug date Figure 9. Daily Popularity for Songs We can see that the popularity of popular songs increase rapidly and reaches the peak after it is released. It needs only about two weeks to get the peak. After that, their popularity decreases continuously, and their life span is similar, which is about 4 to 6 months. rate of listening events.2.18.16.14.12.1.8.6.4.2 Track #1 (Nobody) Track #2 (As I was shot) Track #3 (Day by Day) Track #4 (Gee) Track #5 (Mirotic) Track #6 (Can you hear me?) Track #7 (Love... him) Track #8 (Rain and You) 2 4 6 8 1 12 14 16 18 age of song (elapsed day since released) Figure 1. Daily Popularity over Age of Songs V. CONCLUSION In this paper, we have studied temporal dynamics in users music listening behaviors by analyzing usage log dataset gathered from Bugs Music, which is one of the most popular online music service in Korea. Our usage log dataset contains billions of music listening events for about a year. Before we studied temporal dynamics, we pointed out that there is a difference between the consumption behavior of music and that of others like news, blog, video, etc. Unlike these other contents, people listen to songs repeatedly. Although the number of times a user listens to a particular song shows his preference for that song, service provider should carefully interpret these values since the number of listening events to a particular song is not linearly correlated with his preference. Unlike other contextual dimensions, time dimension has periodicity. We showed that some genres of music are preferred at specific time of day. Moreover, we analyzed seasonal effect and showed that some genres are more popular at particular seasons or months. However, our study classified songs only by genre. Future work should focus on finding which groups of songs are related to which contextual concept using statistical analysis. In addition, for more accurate analysis, we should include a dataset for several years and use more sophisticated techniques. We did a time-series analysis for some popular songs. It shows that popularity of songs changes continuously. After a song is released, its popularity reaches to the peak very fast, then it decreases slowly. Modeling popularity in the context of temporal dynamics and predicting popularity will help maximize users attention and revenue. As we saw in the paper, temporal analysis on usage log dataset reveals users content consuming behaviors. We expect that it can be a useful guideline to other researchers and the industry that aim to deliver online contents to users and consider their personal context. 577

ACKNOWLEDGMENT We would like to thank Neowiz Bugs Corporation who provided their log dataset for research purpose. We are also grateful to Professor Sang-goo Lee, Dongjoo Lee, Sung Eun Park, and Sangkeun Lee at Seoul National University for comments and helping us to co-operate. This research was supported by the MKE(The Ministry of Knowledge Economy), Korea, under the ITRC(Information Technology Research Center) support program supervised by the NIPA(National IT Industry Promotion Agency). (grant number NIPA-21-C19-131-2) REFERENCES [1] F. Wu and B. A. Huberman, Popularity, novelty and attention, in EC 8: Proceedings of the 9th ACM conference on Electronic commerce. ACM, 28, pp. 24 245. [2] Y. Koren, Collaborative filtering with temporal dynamics, in KDD 9: Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 29, pp. 447 456. [3] I. Koychev and I. Schwab, Adaptation to drifting user s interests, in Proceedings of ECML2 Workshop: Machine Learning in New Information Age. Citeseer, 2, pp. 39 46. [4] S. Ma, X. Li, Y. Ding, and M. Orlowska, A recommender system with interest-drifting, in WISE 7: Proceedings of the 8th international conference on Web information systems engineering. Springer, pp. 633 642. [5] D. Agarwal, B. Chen, and P. Elango, Spatio-temporal models for estimating click-through rate, in WWW 9: Proceedings of the 18th international conference on World wide web. ACM, 29, pp. 21 3. [6] M. Vlachos, C. Meek, Z. Vagena, and D. Gunopulos, Identifying similarities, periodicities and bursts for online search queries, in SIGMOD 4: Proceedings of the 24 ACM SIGMOD international conference on Management of data. ACM, 24, pp. 131 142. [7] D. Shin, J. won Lee, J. Yeon, and S. goo Lee, Context-aware recommendation by aggregating user context, in CEC 9: IEEE Conference on Commerce and Enterprise Computing. IEEE, 29, pp. 423 43. [8] O. Alonso, M. Gertz, and R. Baeza-Yates, Clustering and exploring search results using timeline constructions, in CIKM 9: Proceeding of the 18th ACM conference on Information and knowledge management. ACM, 29, pp. 97 16. [9] J. Schafer, D. Frankowski, J. Herlocker, and S. Sen, Collaborative filtering recommender systems, pp. 291 324, 27. [1] Y. Ding and X. Li, Time weight collaborative filtering, in CIKM 5: Proceedings of the 14th ACM international conference on Information and knowledge management. ACM, 25, pp. 485 492. [11] Y. Ding, X. Li, and M. Orlowska, Recency-based collaborative filtering, in ADC 6: Proceedings of the 17th Australasian Database Conference. Australian Computer Society, Inc., 26, p. 17. [12] H. Park, J. Yoo, and S. Cho, A context-aware music recommendation system using fuzzy bayesian networks with utility theory, Fuzzy Systems and Knowledge Discovery, pp. 97 979, 26. [13] E. Pampalk, T. Pohle, and G. Widmer, Dynamic playlist generation based on skipping behavior, in Proceedings of the 6th ISMIR Conference, 25, pp. 634 637. [14] L. Guo, E. Tan, S. Chen, Z. Xiao, and X. Zhang, The stretched exponential distribution of internet media access patterns, in Proceedings of the twenty-seventh ACM symposium on Principles of distributed computing. ACM, 28, pp. 283 294. [15] A. Das, M. Datar, A. Garg, and S. Rajaram, Google news personalization: scalable online collaborative filtering, in Proceedings of the 16th international conference on World Wide Web. ACM, 27, p. 28. [16] G. Adomavicius, R. Sankaranarayanan, S. Sen, and A. Tuzhilin, Incorporating contextual information in recommender systems using a multidimensional approach, ACM Transactions on Information Systems (TOIS), vol. 23, no. 1, pp. 13 145, 25. [17] D. Lee, S. E. Park, M. Kahng, S. Lee, and S. goo Lee, Exploiting contextual information from event logs for personalized recommendation, in 9th IEEE/ACIS International Conference on Computer and Information Science, Studies in Computational Intelligence. Springer, 21. [18] F. Bass, Comments on a new product growth for model consumer durables: the Bass model, Management science, pp. 1833 184, 24. [19] K. Lerman and T. Hogg, Using a model of social dynamics to predict popularity of news, in WWW 1: Proceedings of the 19th international conference on World wide web. ACM, 21, pp. 621 63. 578