<<

Inferring Personal Traits from Music Listening History

Jen-Yu Liu Yi-Hsuan Yang Research Center for IT Innovation Research Center for IT Innovation Academia Sinica, Taipei, Taiwan Academia Sinica, Taipei, Taiwan [email protected] [email protected]

ABSTRACT role in personalized web applications such as personalized Nowadays, we often leave our personal information on the search, recommendation, and advertisement. Internet without noticing it. People could learn things about The way people experience and consume music has also you from these information. It has been reported that it evolved in the digital age. The number of users paying to is possible to infer some personal information from the web subscribe to music services is estimated to have increased browsing records or from blog articles. As the music stream- by 65% in 2011 to 13.4 million worldwide, according to ing services become increasingly popular, the music listening International Federation Phonographic Industry (IFPI) es- history of one person could be acquired easily. This paper timates [1]. Music streaming and Internet radio services such as Pandora, Last.fm, Spotify, Rhapsody, and KKBOX investigates the possibility for a computer to automatically 1 infer personal traits such as gender and age from the music have been increasingly popular. The music listening his- listening history. Specifically, we consider three types of fea- tory records of people could be available either to streaming tures for building the machine learning models, including 1) service companies or to the public. statistics of the listening timestamps, 2) /artist meta- According to psychology studies, when an individual se- data, and 3) song signal features, and evaluate the accuracy lects a particular song, album, or radio station to listen to, of binary age classification and gender classification utiliz- his or her selection is influenced by factors such as emotions, ing a 1K-user dataset obtained from the online music service personal experience, social context, culture, etc [8, 12, 23]. Last.fm. Our study brings about new insights into the hu- In particular, the psychologist Gosling has shown that just man behavior of music listening, but also raises concern over reading the iTunes list of a college student allows other un- the privacy issues involved in music streaming services. related college students to guess the Big-Five personality traits of the target [11]. Given the potential sociological and commercial impact of such a research, it is a bit sur- Categories and Subject Descriptors prising that little work if any has been done in the music H.5.5 [Sound and Music Computing]: Methodologies information retrieval (MIR) community to computationally and techniques, Systems; H.5.1 [Multimedia Information model personal traits from listening profiles. Systems]: Evaluation/methodology In this paper, we are motivated to investigate the possi- bility for a computer to automatically infer personal traits Keywords from the music listening history obtained from an online mu- sic streaming service, which records when and which pieces Personal information retrieval, music recommendation sys- of music a user listened to during a span of time. From prac- tem, listening history, personal traits tical application point of view, being able to infer personal information of a user is helpful for recommending music and 1. INTRODUCTION for providing better service to the user [5,10]. From scientific People might leave a lot of information on the Internet. research point of view, the present study offers insights that For example, web browsing records are available to search can be fundamental to research on user-centered strategies engine companies, and blog articles, when made public, are for MIR applications. available to everyone. It has been reported that personal In particular, this study attempts to answer the following traits such as gender and age could be automatically inferred four research questions. from web browsing records or blog articles by a machine • How effective could the music listening history be used [14, 20, 21]. Such personal information plays an important to infer personal traits such as age and gender?

• What is the importance of temporal patterns of music listening for the above prediction tasks? Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies • Do different groups of people listen to different types bear this notice and the full citation on the first page. To copy otherwise, to of music? republish, to post on servers or to redistribute to lists, requires prior specific 1 permission and/or a fee. http://www.pandora.com/, http://www.last.fm/, MIRUM’12, November 2, 2012, Nara, Japan. http://www.spotify.com/, http://www.rhapsody.com, Copyright 2012 ACM 978-1-4503-1591-3/12/11 ...$15.00. http://www.kkbox.com/

31 • Are the observations in this study consistent with re- lated research on personal traits prediction?

Motivated by Herrera et al. [13], who analyzed the times- tamps of music listening to investigate the listening fre- quency of in a day or a week, we move one step forward and propose several novel time-related features to capture the temporal patterns of music listening, such as the per- centage of songs that are listened to for each hour of a day, each day of a week, or each month of a year. The intuition is that timestamps actually provide rich information about the time a user is engaged in music streaming services. By using them, the daily, weekly, and monthly routine patterns could be deduced. As a consequence, three sets of features are investigated: Figure 1: Data distribution of Last.fm-1K [5]. 1) temporal patterns derived from listening timestamps, 2) song/artist metadata, and 3) song signal features. We also refer to the first and the last two feature sets as context and own self-descriptions, but the simple screen name and full content features, respectively. These features are used to name are obviously more informative. classify data in age and gender classification tasks. The web browsing records have been used for age and gen- The contribution of the paper lies in the proposal of novel der prediction as well [14]. The feature used is simply the context features for music listening as well as the answers frequency of the web pages visited by the user. Experiment to the four research questions that might further our un- results on a collection of page click-through log indicates the derstanding on the user behavior of music listening. In this automatic system can achieve 70.3% on binary gender clas- work we focus on age and gender, for we are not aware of sification and 46.1% on five-class age classification in terms a public dataset for music listening records that come with of macro-averaged f-score. annotations of other personal traits (such as personality, oc- It is noted that in previous works context information such cupation, location, and music background). as the temporal patterns of posting a tweet or browsing the The remainder of the paper is organized as follows. Sec- Internet is seldom exploited. We also note the task of age tion 2 provides a brief review on related work. Section 3 estimation is usually formulated as a classification problem, describes the dataset and the features we have extracted to despite that age is in nature a continuous variable. As one represent the data. Section 4 describes the methodology we of our research problems is to compare our result with the employ for the performance study. Section 5 reports the findings in related fields, we formulate age prediction as a result. We discuss the implications of the evaluation result binary classification problem as well and leave the investi- further in Section 6 and conclude the paper in Section 7. gation of age regression to future work. To the best of our knowledge, this work represents the first attempt to build a computational model for predicting age and gender from 2. RELATED WORK music listening records. There are works that use blog articles to infer age or gen- der. In the paper [20], the authors acquired chat texts from a 3. CORPUS AND FEATURES social networking site Netlog and performed several types of The listening history for users of the famous music stream- classification tasks, including three sets of binary age classifi- ing service Last.fm can be obtained by the Last.fm API.2 For cation tasks: {younger than 15 vs. older than 16}, {younger reproducibility, in this study we utilize the publicly-available than 15 vs. older than 18}, {younger than 15 vs. older than Last.fm-1K users dataset3 [5]. It contains listening history 25}. Features such as uni-grams and bi-grams are extracted records of 992 users over years collected by using the afore- from the text to represent the data, which means the clas- mentioned API. There are a total of 19,150,868 lines of infor- sifier needs to distinguish age from the choice of words or mation in the dataset. Each line contains the user ID, song phrases utilized in online chat. The support vector machine title, artist name, and timestamps (in the format of yyyy- (SVM) based classifier showed promising results for classi- mm-dd hh:mm:ss). Most of the music are Western popular fying adults versus adolescents and yielded an accuracy of music. For each user, the information of age, gender, and 71.3% for the first classification task. As the distance be- nationality is collected, but there are missing values. Among tween the two age groups increased, the accuracy rose to the 992 users, 286 users have age information and 884 users 80.8% and 88.2% for the last two tasks, respectively. have gender information. See Figure 1 for the age and gen- Twitter texts have also been utilized for gender classifica- der distributions of the dataset. In this study, we choose not tion [4]. Using a collection of gender-tagged tweets, Burger to work on nationality because the data is biased to Western et al. explored several types of text-based or metadata-based people. features as input to the machine learning model, including It should be noted that the dataset might be biased to “all tweet texts,” “self-description” (e.g., retired accountant tech-savvy people because one needs to have a Last.fm ac- and grandfather), “screen name” (e.g., jsmith86), and “full count to subscribe to the service. The result we obtain from name” (e.g., Joe Smith). The accuracy for gender classifica- this study may not be applicable to people who prefer to lis- tion obtained by using different feature sets is 75.5%, 71.2%, 77.1%, and 89.1%, respectively. It appears that the tweet 2http://www.last.fm/api texts convey more about a Twitter user’s gender than their 3http://ocelma.net/MusicRecommendationDataset/

32 ten to music they downloaded or purchased instead of music artist histogram, whose dimension is close to that of the that is streamed over the Internet. song histogram (i.e., 10,000). By using Last.fm API, the 100 top tags and the tag weights 3.1 Context Features: Temporal Patterns for each song and each artist can be obtained. Then, we se- We propose to derive context features from the times- lected 10,000 tags that have the largest total weights. For tamps of the listening records. People in different groups each user, a 10,000-D song tag histogram is computed by might have different routine patterns, which effect the time taking the weighted sum of the song tags according to the available for using online music services. For example, it user’s listening song counts. Similarly, a 10,000-D artist has been reported that there are gender difference in rou- tag histogram is computed. tine patterns [2]. However, how such difference influences music listening remains unclear. 3.2.2 Song Signal Features To investigate the importance of temporal patterns of mu- We use the Echo Nest’s public API for extracting song sic listening for the prediction of personal traits, the follow- signal features such as , pitch, and timbre [3]. Many ing features are computed from the listening history of each of these features have been used extensively in MIR research user. [7, 22, 24, 28]. • Hour-of-day histogram: 24-D histogram that sim- • Danceability: The ease with which a person could ply counts the number of songs a user listens to in each dance to a song, over the course of the whole song [22]. hour of a day. Normalize so that it sums to one. • Loudness: Perceived intensity of a sound. • Hour-of-day entropy: Entropy of the hour-of-day histogram. • Key: The estimated overall key of a track (ranges from 0 to 11 and corresponds to one of the 12 keys: C, • Working-hour ratio: Percentage of songs that are C#, D to B). listened to from 8 to 19 o’clock in a day. • Mode: The estimated mode; 0 (minor) or 1 (major). • Day-of-week histogram: 7-D histogram that simply counts the number of songs a user listens to in each day • Tempo: The overall estimated tempo of a track in of a week. Normalize so that it sums to one. beats per minute (BPM). • Day-of-week entropy. • Pitch: Frame-by-frame chromavector, corresponding to the 12 pitch classes [24]. We take the mean and • Working-day ratio: Percentage of songs that are lis- standard deviation (SD) for temporal integration [16]. tened to from Monday to Friday in a week. • Timbre: Frame-by-frame MFCC-like feature vector, • Month-of-year histogram: 12-D histogram that sim- corresponding to the 12 basis functions that are loosely ply counts the number of songs a user listens to in each related to perceptual qualities such as brightness, flat- month of a year. Normalize so that it sums to one. ness, and attack strength [7]. We also take the mean • Month-of-year entropy. and SD for temporal integration. • Working-month ratio: Percentage of songs that are As it is infeasible to extract the song signal features for all listened to in the “working months” (excluding July the songs, we only extract the 15 most frequently listened and August) in a year. songs for each user for feature extraction. For each feature, the weighted-average is calculated and taken as the repre- It can be found that the first three, middle three, and last sentative value of this feature for the user. The weights are three features are in different temporal scales: the hours set according to the normalized listened counts of these 15 of a day, the days of a week, and the months of a year, songs with sum-to-one normalization. respectively. The histogram-related features (e.g., month- of-year histogram) are normalized by sum-to-one so that the histograms of different users can be compared, regardless of 4. METHODOLOGY the difference in the total number of songs a user listens to. Two binary classification tasks are performed: gender clas- sification (female or male) and age classification (under or 3.2 Content Features above 24 year-old). The continuous-valued attribute age is In contrast to the context features, the content features discretized to ‘adolescents’ and ‘adults’ at the cutoff point are derived from the songs. Two types of content features are 24, mainly because 24 year-old happens to be the median extracted: song/artist metadata and song signal features. of the dataset (cf. Figure 1). This practice has been uti- lized in previous work for avoiding data sparsity of either 3.2.1 Song/Artist Metadata class [19]. For age classification, there are 144 adolescents Four features are computed: song/artist histogram and and 142 adults. For the gender classification, there are 382 song/artist tag histogram. A song histogram is an M- females and 502 males. In order to have a balanced data, bin histogram that counts how often a song (among the M we randomly select 382 males and keep all the females. possible songs) is listened to by a user. For computational SVM is utilized to train the classifiers for its superior per- efficiency, we select the 50 most frequently listened songs formance demonstrated in a large variety of machine learn- from each of the 992 users, and then select the top 10,000 ing problems [6]. SVM is a supervised classification method most frequently listened songs (from the resulting 49,600 that attempts to draw a decision boundary between classes songs) to form the song histogram. In addition, we group so as to maximize the gap between them. The use of a ker- the 49,600 songs by artist names and calculate the artist nel nonlinearly projects data to a high dimensional space, histogram for each user. This gives rise to a 9,467-bin where data are believed to be more separable. Our imple-

33 Feature Age Gender Feature Age Gender Hour-of-day histogram 55.7% 57.0% Artist histogram 71.1% 65.8% Hour-of-day entropy 45.7% 57.1% Artist tag histogram 60.0% 62.2% Working-hour ratio 47.5% 48.4% Song histogram 64.6% 66.1% Day-of-week histogram 58.9% 47.2% Song tag histogram 58.9% 63.6% Day-of-week entropy 61.4% 48.9% Danceability 46.4% 52.2% Working-day ratio 61.1% 47.0% Loudness 50.4% 49.7% Month-of-year histogram 50.4% 47.5% Key 50.4% 46.6% Month-of-year entropy 49.3% 50.4% Mode 52.1% 52.8% Working-month ratio 50.0% 50.4% Tempo 46.4% 50% Pitch 52.9% 54.3% Table 1: Classification accuracy of context features Timbre 59.3% 53.7%

Table 2: Classification accuracy of content features

further shows the average day-of-week histogram for ado- lescents and adults, with each error bar representing one standard error of all the users belong to that group. It can be found that the distribution is fairly even, but adults tend to use Last.fm less during the weekends. We consider this as a media-dependent phenomenon: the dataset used in this study is from the online music service Last.fm, so the users might have to be in front of a computer Figure 2: Average day-of-week histogram for to access the service. In consequence, adults might have younger users (adolescents) and older users (adults). more chances to use the services when they have desk jobs The cutoff age between adolescents and adults is 24. during the weekdays. During the weekends, some of the adults might prefer other activities, while some have more free time listening to music. This may help explain the larger mentation of SVM is based on the LIBSVM library.4 We use standard deviation for adults on Sundays. the radial basis function (RBF) kernel and perform five-fold cross-validation on the training data to choose the param- 5.1.2 Gender Classification eters C and γ. Each task is performed five times and take Similar to age classification, the best accuracy achieved average, where the training data and testing are randomly for gender classification, 57.1%, is not that high. It can be chosen according to the training rate each time. The train- observed that the highest results are achieved by hour-of- ing rate is set to 0.8. day histogram and hour-of-day entropy, which are both at the hour-of-day scale. The performance difference between the best result and that of random guessing is significant 5. RESULT under the two-tailed t-test at 1% significance level. The result of age and classification using different features In [2], Adan et al. studied the morningness-eveningness are shown in Tables 1 and 2. Values in bold font represent preference of 1,041 male and 1,094 female college students, significant ones, comparing to a random guessing baseline. who were asked to answer the morningness-eveningness ques- Below we discuss the results in detail. tionnaire (MEQ). According to MEQ, people are designated to morning-type, evening-type, or neither-type. For male, 5.1 Result of Context Features 13.6% are morning-type, 28.8% are evening-type, and 57.6% are neither-type. For female, the percentages are 17.9%, 5.1.1 Age Classification 20.5%, and 61.6%, respectively. Males are more likely to be The best accuracy achieved by context features is 61.4%, evening-type compared to females, while females are more which shows that there is certain correlation between tem- likely to be morning-type. From Figure 3, it can be seen poral patterns of music listening with age, but not that that females listened to more songs between 17–22 o’clock overwhelming. More interestingly, from Table 1 we could than males, whereas males listened to more songs between observe that the top three performing features for age clas- 6–12 o’clock. Does this result contradicts the findings in [2]? sification are achieved by the features at the day-of-week The following observations may help answer the above scale. It might be thought that the adolescents can listen question. First of all, the Last.fm-1k dataset was collected in to more songs on weekdays than adults because adolescents 2009, when the smart phones were not as common as nowa- have more free time than older ones during the weekdays, days. Therefore, it might be reasonable to assume that most and this difference in turn leads to the pattern difference at people were near a computer when they used the Last.fm ser- the day-of-week scale. However, after a closer look, we found vice. In addition, music listening does not require our full that the average of working-day ratio for adults (0.733) is attention, so people can listen to music almost in any cir- larger than that for adolescents (0.706), which can be con- cumstance as long as they are near a computer. The MEQ firmed by t-test under the 5% significance level. Figure 2 measures the preference of sleep time, the time when sub- jects feel more energetic, the time when the subjects want to 4http://www.csie.ntu.edu.tw/~cjlin/libsvm do physical exercise, etc. In other words, it measures when

34 Figure 3: Average hour-of-day histogram for male and female users. The vertical axis represents the average percentage of songs that are played during a specific hour of a day. the subjects are more active. When someone is less active, 6. DISCUSSION s/he can still listen to music. On the contrary, an energetic Generally speaking, our evaluation shows that music lis- individual might like go exercising, shopping, hanging out tening history can also be utilized to infer information about with friends, etc. Therefore, if music listening is considered age and gender, and that many findings in related fields as an activity that is “less active,” what we observe in Figure still hold of music. Compared to the age classification by 3 is consistent with the findings reported in [2]. blog articles [20] described in Section 2, the age classifica- tion accuracy (71.1%) by music listening records is com- parable with the accuracy in their first classification task 5.2 Result of Content Features (71.3%). As for the gender classification task performed by tweet texts [4] described in Section 2, the “all tweet texts” By comparing Tables 1 and 2, it can be found that con- feature is more similar to the music listening history we use tent features generally outperform context features, espe- in this study, because both of them are publicly available cially for song/artist metadata. Among the content features, and contain many pieces of information across time. The the artist histogram and the song histogram lead to the best best accuracy achieved by music listening records is 66.1%, result 71.1% and 66.1% for age classification and gender clas- which is slightly inferior to the performance of “all tweet sification, respectively. Both result are significantly better texts” (71.2%) for gender classification in text. than the result obtained with context features. It can be In addition to the classification methods and the accu- seen that the artist and song histograms perform well for racy achieved, we should also pay attention to the possible both age and gender classification. privacy issues that this kind of data and researches could In [23], Schwartz and Fouts pointed out that there are raise. Two famous cases are the “AOL search data leak”5 gender differences in music preferences. For example, female and the “Netflix movie-picking contest”6 privacy issue. For adolescents prefer light music more than male adolescents example, to organize a movie recommendation competition, do, whereas male adolescents prefer heavy music qualities Netflix released a large collection of user movie-rating data, more than female adolescents do. The effectiveness of the with the name of the raters anonymized. However, it has song and artist histograms is consistent with these findings. been reported that some private information could still be However, among the song signal features we have consid- uncovered by some computational techniques [18]. In fact, ered, only the timbre features could be considered as effec- the music listening records are similar to movie-rating data tive for the task of age classification. The accuracy in age in several aspects. The play counts on songs and artists classification for timbre features is 59.3%, which is slightly are analogous to the ratings of movies. The music listen- inferior to the context features at the day-of-weak level. ing records might reveal more details about personal daily For gender classification, the results of all the song signal schedule, as music can be listened to most of the time. For features are not significantly better than random guessing. the current state of our study, we are only able to retrieve Generally speaking, the performance of song signal features information about gender and age to a certain extent. But is even inferior to that of the context features, possibly be- in the future, more information could be retrieved and re- cause the way we aggregate temporal information by taking searchers in this area should aware of this kind of issues. mean and STD is not effective (cf. Section 3.2.2), or because the song signal features cannot capture the nuance for music preference in different groups. 7. CONCLUSION To offer a qualitative evaluation, in Table 3 we show the In this paper, we have presented a study that attempts most listened songs of the Last.fm-1K dataset for different to automatically infer gender and age information from a user groups. It can be found that the selection of songs is collection of music listening history. Our main findings are indeed different for people in different groups. Interestingly, as follows. First, the age and gender information could be for this particular dataset, there is a common preference in estimated with certain accuracy by the song listening his- the songs of for male listeners and adolescent listeners. In the adolescent group, there are 74 males and 5http://en.wikipedia.org/wiki/AOL_search_data_leak 65 females, so the common preference for this singer among 6http://www.usatoday.com/tech/news/2010-03-12- adolescent listeners should not result from the gender bias. netflix-contest-privacy_N.htm

35 Song title and artist name Count tags. In Proc. ACM Int. Conf. Multimedia Information Female Retrieval, pages 55–61, 2010. Wings of Words by Chemistry 3112 [8] P. G. Dunn, B. de Ruyter, and D. G. Bouwhuis. Toward a Heartbeats by the Knife 2632 better understanding of the relation between music Street Lights by Kanye West 2006 preference, listening behavior, and personality. Psychology of Music, pages 1–18, 2011. Staring at the Sun by TV on the Radio 1891 [9] N. Eagle and A. S. Pentland. Eigenbehaviors: identifying Such Great Heights by the Postal Service 1855 structure in routine. Behavioral Ecology and Sociobiology, Male 63(7):1057–1066, 2009. The Good, the Bad, the Queen 3149 [10] H. Garcia-Molina, G. Koutrika, and A. Parameswaran. Gimme More by Britney Spears 3111 Information seeking: Convergence of search, by Kanye West 2812 recommendations and advertising. Communications of the Heartless by Kanye West 2492 ACM. Accepted for publication. Snoop: What Your Stuff Says About You [11] S. Gosling. . Basic by Kanye West 2357 Books, 2008. Adolescent [12] S. Hallam et al. The Oxford Handbook of Music Wings of Words by Chemistry 3112 Psychology. Oxford University Press, New York, 2008. Love Lockdown by Kanye West 2265 [13] P. Herrera, Z. Resa, and M. Sordo. Rocking around the Heartless by Kanye West 2211 clock eight days a week: An exploration of temporal Paranoid by Kanye West 2136 patterns of music listening. In Proc. ACM Workshop on Amazing by Kanye West 2122 Music Recommendation and Discovery, 2010. [14] J. Hu, H.-J. Zeng, H. Li, C. Niu, and Z. Chen. Adult Demographic prediction based on user’s browsing behavior. Hung Up byMadonna 927 In Proc. Int. Conf. World Wide Web, pages 151–160, 2007. Bright Lights by Matchbox Twenty 894 [15] V. Laube, C. Moewes, and S. Stober. Browsing music by Better Alone by 835 usage context. In Proc. Workshop on Learning the Downfall by Matchbox Twenty 743 Semantics of Audio Signals, pages 19–29, 2008. Enjoy the Silence by Depeche Mode 701 [16] A. Meng et al. Temporal feature integration for classification. IEEE Trans. Audio, Speech, and Language Processing Table 3: Top listened songs for different groups , 15(5):1654–1663, 2007. [17] A. Nanopoulos et al. MusicBox: Personalized music recommendation based on cubic analysis of social tags. IEEE Trans. Audio, Speech, and Language Processing tory, despite that the prediction accuracy is not as good , 18:407–412, 2010. as that from other media such as web browsing records or [18] A. Narayanan and V. Shmatikov. Robust de-anonymization blog articles. Second, the importance of time scales is task- of large sparse datasets. In Proc. IEEE Symp. Security and dependent. For age classification, the day-of-week scale is Privacy, pages 111–125, 2008. more effective. In contrast, for gender classification, the [19] O. Omojokun, M. Genovese, and C. Isbell. Impact of user hour-of-day scale is more effective. Third, artist and song context on song selection. In Proc. ACM Int. Conf. metadata perform the best among the features we consid- Multimedia, pages 1057–1066, 2008. ered, followed by time-related features. Finally, the obser- [20] C. Peersman, W. Daelemans, and L. Van Vaerenbergh. Predicting age and gender in online social networks. In vations in this paper are also consistent with the findings Proc. Int. Workshop on Search and Mining in related fields. For future work, we are interested in in- User-Generated Contents, pages 37–44, 2011. corporating other context factors such as location [9] and [21] S. Rosenthal and K. McKeown. Age prediction in blogs: A environment [19], or devise more sophisticated features that study of style, content, and online behavior in pre- and are potentially useful. post-social media generations. In Proc. Annual Meeting of the Association for Computational Linguistics, pages 8. ACKNOWLEDGMENTS 763–772, 2011. [22] B. Schuller, F. Eyben, and G. Rigoll. Tango or Waltz? This work was supported by the National Science Council putting ballroom dance style into tempo detection. of Taiwan under contracts NSC-100-2218-E-001-009, NSC EURASIP J. Audio, Speech, and Music Processing, 2008. 101-2221-E-001-017, and a grant from HTC. Article ID 846135, 12 pages. [23] K. D. Schwartz and G. T. Fouts. Music preferences, 9. REFERENCES personality style, and developmental issues of adolescents. [1] Digital Music Report, 2012. [Online] J. Youth and Adolescence, 32:205–213, 2003. http://www.ifpi.org/content/library/DMR2012.pdf. [24] J. Serr`a,E. G´omez,P. Herrera, and X. Serra. Chroma [2] A. Adan and V. Natale. Gender differences in binary similarity and local alignment applied to cover song morningness-eveningness preference. Chronobiology International: The J. Biological & Medical Rhythm identification. IEEE Trans. Audio, Speech, and Language Processing, 16:1138–1151, 2008. Research, 19(4):709, 2002. [25] M. Slaney. Web-scale multimedia analysis: Does content [3] T. Bertin-Mahieux et al. The million song dataset. In Proc. matter? IEEE Multimedia, 18(2):12–15, 2011. Int. Society of Music Information Retrieval Conf., 2011. [26] A. Uitdenbogerd and R. van Schnydel. A review of factors [4] J. D. Burger et al. Discriminating gender on twitter. In Proc. Int. Society Proc. Int. Conf. Empirical Methods in Natural Language affecting music recommender success. In of Music Information Retrieval Conf. Processing, pages 1301–1309, 2011. , pages 204–208, 2002. [5] O. Celma. Music Recommendation and Discovery in the [27] F. Weninger, M. Wollmer, and B. Schuller. Automatic assessment of singer traits in popular music: Gender, age, Long Tail. Springer, 2010. height and race. In A. Klapuri and C. Leider, editors, [6] C. Cortes and V. Vapnik. Support vector networks. ISMIR, pages 37–42. University of Miami, 2011. Machine Learning, 20:273–297, 1995. [28] Y.-H. Yang and H. H. Chen. Music Emotion Recognition. [7] D. T. Derek Tingle, Youngmoo E. Kim. Exploring CRC Press, 2011. automatic music annotation with acoustically-objective

36