A unified music recommender system using listening ...

2 downloads 223 Views 493KB Size Report
both listening habits and semantics of tags in a social music site. Most ... big data analytics, semantic technologies, and recommender systems. Donggeon Kim is ...
14

Int. J. Intelligent Information and Database Systems, Vol. 8, No. 1, 2014

A unified music recommender system using listening habits and semantics of tags Hyon Hee Kim*, Donggeon Kim and Jinnam Jo Department of Statistics and Information Science, Dongduk Women’s University, 60 Hwarang-ro 13-gil Seongbuk-Gu, Seoul, Korea E-mail: [email protected] E-mail: [email protected] E-mail: [email protected] *Corresponding author Abstract: In this paper, we propose a unified music recommender system using both listening habits and semantics of tags in a social music site. Most commercial music recommender systems recommend music items based on the number of plays, or explicit ratings of a song. However, these approaches have some difficulties in recommending new items with only a few ratings, or recommending items to new users with little information. To resolve the problem, UniTag ontology is developed, which defines the meaning and the weighted score of tags. User profiles are created by combining the score of tags and the number of plays, and a collaborative filtering algorithm is executed. For performance evaluation, precisions, recalls, and F-measures are measured using the listening habits-based recommendation, the tag score-based recommendation, and the unified recommendation, respectively. Our experiments show that the proposed approach outperforms the other two approaches in terms of all the evaluation metrics. Keywords: music recommendation; tag ontology; intelligent information and database systems; emotion ontology; listening habits. Reference to this paper should be made as follows: Kim, H.H., Kim, D. and Jo, J. (2014) ‘A unified music recommender system using listening habits and semantics of tags’, Int. J. Intelligent Information and Database Systems, Vol. 8, No. 1, pp.14–30. Biographical notes: Hyon Hee Kim is an Assistant Professor at the Department of Statistics and Information Science, Dongduk Women’s University, South Korea. She received her MS and PhD in Computer Science from Ewha Womans University, South Korea. Her current research interests are big data analytics, semantic technologies, and recommender systems. Donggeon Kim is a Professor at the Department of Statistics and Information Science, Dongduk Women’s University, South Korea. He received his MS and PhD in Statistics from Virginia Polytechnic Institute and State University. His current research interests are data mining and statistical analysis. Jinnam Jo is a Professor at the Department of Statistics and Information Science, Dongduk Women’s University, South Korea. He received his PhD in Statistics from Virginia Polytechnic Institute and State University. His current research interests are design of experiments and sampling design and analysis. Copyright © 2014 Inderscience Enterprises Ltd.

A unified music recommender system using listening habits

15

This article is a revised and expanded version of a paper entitled ‘Generation of tag-based user profiles for clustering users in a social music site’ presented at The 4th Asian Conference on Intelligent Information and Database Systems, Kaoshiung, March 2012.

1

Introduction

With recent progress in Internet technology, the music industry changed dramatically. People listen to music or buy digital music items through a social music site such as last.fm. Also, they make friends with other users with similar musical tastes, and insert keywords that they select to music items, using a collaborative tagging system. Since most music sites have a large number of music items, it is essential to recommend music items to users. There is consideration for a music recommendation system comparable to those of other products, such as books or movies. In the case of books or movies, after buying products users rate them with an explicit rating system, whereas in the case of music, users tend to listen to music repeatedly instead of rating the music item. That is, implicit information, like the number of plays, is more important to catch user preferences. Therefore, listening habits are generally used for user profiles in a music recommender system. The generation of user profiles based on listening habits causes two main problems. One is the well-known cold start problem. New items with a few ratings are rarely recommended. Also, it is difficult for new users with little information to recommend music items. The other is the data sparsity problem. The number of music items that users have listened to is very much smaller than the number of music items that need accurate predictions. The aim of this paper is to propose a novel music recommender system for a social music site. To resolve the conventional problems mentioned above, tag information generated by collaborative tagging is used. Collaborative tagging is a functionality enabling users to add keywords chosen by themselves to web resources. It has becomes popular in most social networking sites. Users add tags to categorise items in their own way, or use tags that other users have added, to browse and search items. In the case of music sites, the role of tags is more important than for other social networking sites. Since about 70% of tags are musical genre or musician, they directly represent users’ musical preferences. In addition, about 10% of tags are emotional tags, which contain positive or negative opinion of the music items, and they also contain users’ musical preferences (Lamere and Pampalk, 2008). Therefore, recently it has becomes popular to do research on music recommender systems using tag information (Nanopoulos et al., 2010). Since the generation of user profiles based on tag information considers the semantics of tags, the predictive accuracy can be improved, compared to the generation of user profiles based on listening habits (Kim, 2012). However, in this case, predictive sensitivity is somewhat low, because users can add tags to music items without actually listening to them. To keep high accuracy as well as sensitivity, a unified approach,

16

H.H. Kim et al.

generating user profiles by combining tag information and users’ listening habits, might be a good solution. The contribution of this paper is in the development and validation of the unified recommender system that considers both listening habits and the semantics of tags. Our recommender system adopts the ontology technology that has been established for processing the semantics of data (Gruber, 1995). Gruber (2007) formed the basis of a tagging ontology that defines relationships among web resources, users, and tags as a tripartite relation, with an object, a tag and a tagger. We developed a tagging ontology, called UniTag to define the meaning of tags, to resolve the semantic ambiguity of tags, and to assign weighted values to the tags. UniTag ontology is composed of UniMusic ontology and UniEmotion ontology. UniMusic ontology, which is an extension of the music ontology (Raimond et al., 2007), defines not only the terms of musical work, genre, and artist, but also rules for resolving synonym, homonym, and acronym. UniEmotion ontology classifies emotional tags as positive tags and negative tags and assigns weighted values to them, according to the intensity of emotion. We implemented a prototype music recommender system by combining listening habits and tag information. A dataset of 1,000 users, tags that they added, and the music items that they listened to, was collected from last.fm. To compare our approach with other approaches, a listening habits-based recommender system and a tag-based recommender system were also implemented. For the performance evaluation, precision, recall, and F-measure were measured respectively. Statistical validation was performed using one-way ANOVA at a 99% confidence level for the three approaches. Our experiments show that the proposed approach significantly improves performance in terms of precision, recall, and F-measure. The remainder of this paper is organised as follows: In Section 2, we discuss related work, and In Section 3, we give an overview of a unified music recommendation system. Section 4 describes the method of unified music recommendation in detail, and in Section 5, we explains how to validate our prototype system. Section 6 shows the experimental results, and finally, Section 7 gives concluding remarks.

2

Related work

In this section, we briefly present some of the research literature related to tag-based recommender systems and music recommendations. We also present research on tag ontology. Recent studies have focused on exploiting tag information as user profiles. Firan et al. (2007) proposed tag-based profiles, which collect tags together with corresponding scores representing the users’ interests as well as track-based profiles. Their experimental results show that tag-based music recommendation improves result’s quality significantly. Durao and Dolog (2010) provide personalised recommendations by evaluation of basic similarity between tags. Further, they extend the calculus with additional factors, such as tag popularity, tag representativeness, and affinity between users and tags. To extract semantic relatedness between two tags, they used WordNet dictionary and ontologies from open linked data. To improve tag-based recommendations, the classification of tags into content-based, context-based, subjective, and organisational categories has been done (Cantador et al.,

A unified music recommender system using listening habits

17

2011). To remove semantic ambiguity, such as acronym, misspelling, or compound nouns, each tag is preprocessed, and then mapped to YAGO ontology, which is an ontological knowledge bases containing information from WordNet and Wikipedia (Suchanek et al., 2008). Once a mapping is found, the YAGO ontology finds a subcategory of tags, and a recommendation algorithm considering tags is executed. Cantador et al. show that the tag-based recommendation improves recommendation performance. In particular, tags play a central role in music recommendation (Lamere and Pampalk, 2008). MusicBox is a personalised music recommender system based on social tags (Nanopoloulos et al., 2010). To capture the 3-way correlations between users, tags, and music items, a 3-order tensors model is used. Nanopoulos et al. show that the proposed method improves the recommendation quality. Since tags are added in the users’ own way, there is no semantic agreement. Therefore, it is necessary to define semantic concepts of the tags, and relationships among them. From this point of view, research on tag ontology has attracted attention. Gruber formed the basis of a tag ontology, which is represented by an object, a tag, and a tagger (Gruber, 2007). Recent research on tag ontology has focused on definition of the tags, and resolving semantic ambiguity of tags (Kim et al., 2008). However, research on emotional tags, which represent users’ emotion, opinion, or sentiment, has rarely been considered. Since the emotional tags contain directly users’ preferences, using the emotional tags for recommender systems is necessary. OntoEmotions (Baldoni et al., 2012) is an ontology of emotional categories covering the basic emotions: Sadness, Happiness, Surprise, Fear, and Anger. Arsmeteo, the art portal, allows a user to add tags and to retrieve artworks via the tags. A tag in Arsmeteo folksonomy is mapped to a concept of OntoEmotions, and new relations can be inferred by reasoning on the ontology of emotions. However, to our best knowledge, there is no available work that applies emotional tags to recommender systems.

3

Overview of the system

Figure 1 illustrates the basic architecture of the unified music recommender system. The system is largely composed of five components: the UniTag ontology, tag score generator, listening habits collector, user profile generator, and music recommender engine. The listening habits collector, beside the user databases, extracts users, music items that the users listened to, and number of plays of the music items. The tag score generator on the left side of Figure 1 extracts users, tags that the users add, and music items that the users add the tags to, from user databases. Tags are processed using UniTag ontology. UniTag ontology is composed of UniMusic ontology, which is an extension of the standard music ontology, and UniEmotion, which refers to SentiWordNet. UniMusic ontology resolves semantic ambiguity of factual tags, and UniEmotion ontology assigns weighted values to emotional tags. The tag score generator sends the score of tags to the user profile generator. The user profile generator in the centre of Figure 1 creates user profiles combining the number of plays from the listening habits collector and the score of tags from the tag

18

H.H. Kim et al.

score generator. The user profile generator sends the unified user profiles to the music recommender engine on the right side of Figure 1. The music recommender engine recommends music items to users. It uses the conventional collaborative filtering algorithm. Figure 1

4

Overview of the system

The unified music recommendation

The unified music recommendation goes through three steps. The first step is tag processing, and the second step is generating user profiles. Finally, the third step is executing the music recommendation algorithm. In Section 4.1, we describe in detail how the semantics of tags are extracted using ontology. In Section 4.2, we present the generation of three types of user profiles, and in Section 4.3, we explain the music recommendation algorithm.

4.1 Extracting semantics from tags using ontology There are three ontologies in this research. The UniTag ontology is a tag ontology, which is modeled as a tripartite relation composed of users, tags, and items (Gruber, 2007). The UniMusic ontology is a domain ontology, describing music items. Finally, the UniEmotion ontology is an emotion ontology, to define emotional tags. Let us take a closer look at the three ontologies in detail.

A unified music recommender system using listening habits

19

4.1.1 The UniTag ontology For the UniTag ontology, we adopt the basic concept of the tag ontology, and augment it with reasoning rules for resolving semantic ambiguity. Therefore, the UniTag ontology cooperates with the UniMusic ontology to resolve semantic ambiguity of tags, and with the UniEmotion ontology to assign a weighted value to a tag. The UniTag ontology has three basic classes: UniTag:Users, UniTag:Items, and UniTag:Tags. •

UniTag:Users explains users in a social music site. It has three properties: UniTag:userID, UniTag:hasAdded, and UniTag:hasAddedTo. UniTag:userID is a unique identifier assigned to each user. UniTag:hasAdded property represents a tag identifier that a user has added. UniTag:hasAddedTo represents an item identifier that a user has added a tag to.



UniTag:Items explains items in a social music site. It has a UniTag:itemID property, which is a unique identifier assigned to each items.



UniTag:Tags describes tags, and has two properties: UniTag:tagID, UniTag:tagName. UniTag:tagID is a unique identifier assigned to each tag, and UniTag:tagName is a real value of a tag. Here, standardised tags are defined in this class. A standardised tag is a representative tag that represents diverse tags containing the same meaning. For example, hip-hop is the standard tag of the tags such as hip-hop, hip hop, etc.

4.1.2 The UniMusic ontology Next, the UniMusic ontology is used for the tag standardisation. It is an extension of the music ontology (Raimond et al., 2007), which provides a standard base for musical information. The music ontology describes music information at three levels of details: Level 1 describes top-level editorial information, level 2 describes the process behind the production of music, and finally level 3 describes the structure and component events of the music being played. The UniMusic ontology includes level 1 description, and defines reasoning rules for the tag standardisation. The major classes and their properties of the UniMusic ontology are explained as follows: •

UniMusic:Work explains a work of music. It has a unique identifier and describes the title with UniMusic:workID and UniMusic:hasTitle properties, respectively. Also, the class has a UniMusic:yearofWork property, to explain the year in which a musical work was done. A work is performed by artists, and the relationship is represented by UniMusic:PerformedBy property.



UniMusic:Genre describes a trend of music according to historical time. It has three subclasses, UniMusic:Standard, UniMusic:Prefix, and UniMusic:Variation. UniMusic:Genre class defines each genre tag that users added, while UniMusic:Standard represents a standardised musical genre without semantic ambiguity. UniMusic:Variation class defines various representations of the genre. A genre tag might have a prefix, which represents country, types of specific musical genre, etc. The prefixes are defined in the UniMusic:prefix class. The UniMusic:Genre class has a property UniMusic:mappedTo. The property assigns the current genre tag to the standardised tag.

20 •

H.H. Kim et al. UniMusic:Artist describes a person or a group of people who performs a musical work. It has two properties: UniMusic:artistID and UniMusic:hasName. UniMusic:artistID specifies the unique identifier of an artist, and UniMusic:hasName explains the name of the music artist.

For the tag standardisation, two types of rules are defined as follows: rules for reasoning prefixes of tags and rules for resolving synonym.

Rules for reasoning prefix After analysing a set of tags in last.fm, we realised that there exist three types of prefixes ahead of the genre tag. The first is a prefix for nationality, which is often used to represent musical style of a genre according to nationality. For example, French rock, English rock, or Spanish rock is each a kind of rock. The second is a prefix for combining two musical genres, such as folk rock, punk rock, or progressive rock. Each of those rocks is a kind of rock. The third is a prefix for general use of a musical genre, such as post rock. Therefore, all of those tags should be mapped to a standard tag, rock. Rule 1 If the prefix of a tag is an instance of UniMusic:Prefix class and the remaining term of the tag is an instance of UniMusic:Standard class, then the tag is mapped to the instance of UniMusic:Standard. (Tag(?t) ∧ Prefix(?p) ∧ tagPrefix(?t, ?p) ∧ subTag(?t, ?s) ∧ Standard(?s) → mappedTo(?t, ?s)

We choose 10 representative musical genres, i.e., rock, hip-hop, electronic, metal, jazz, rap, funk, folk, blues, and reggae in the order of most frequently used tags. Therefore, the rules cover only the selected ten representative genres.

Rules for resolving a synonym The rules map diverse expressions of a musical genre to a standard musical genre. Several variations of a musical genre are defined in advance, and then those variations are mapped to the standard musical genre. If a tag is a variation of a standard tag, and the other tag is the same as the former tag, then the later tag is also defined as the same standard tag. Rule 2 If a tag is an instance of UniMusic:Variation, then the tag is mapped to an instance of UniMusic:Standard. (Tag(?t) ∧ Variation(?t, ?v) ∧ Standard(?v) ∧ Tag(?s) ∧ SameAs(?t, ?s) → mappedTo(?s, ?v)

For the UniMusic:Variation class, foreign language and synonym is defined for the 10 representative musical genres. For example, rock and roll is defined as the synonym of rock, and hip-hop and hip hop are defined as the synonym of hip-hop. Also, rhythm and blue and R&B are the synonym of blues.

A unified music recommender system using listening habits

21

4.1.3 The UniEmotion ontology Finally, the UniEmotion ontology defines emotional tags in a social music site, and assigns the weighted values to the tags according to the intensity. In the case of last.fm, 85% of tags are factual tags related to genre, region, and instrumentation, while 10% of tags are emotional tags related to opinion such as opinion or sentiment. The remaining 5% of tags are personal tags or organisational tags. Although the emotional tags occupy 10% of tags, we emphasise that the role of the emotional tags in the music recommender system is crucial. A factual tag, e.g., progressive rock and an emotional tag, e.g., perfect should not be considered to have the same weight. Therefore, the UniEmotion ontology classifies emotional tags into positive tags and negative tags, and assigns weights to them. Figure 2 shows classification of emotional tags in UniEmotion. Figure 2

Classification of emotional tags in UniEmotion (see online version for colours)

The UniEmotion ontology is composed of four negative emotions, i.e., fear, sadness, disgust, anger, and four positive emotions, i.e., joy, trust, anticipation, surprise chosen by Plutchik’s model (1997). Intensities of an emotion are defined as strong, middle, and weak. In Figure 2, as a circle becomes bigger, the intensity of emotion becomes weaker. To define the intensity of an emotional tag, SentiWordNet (Esuli and Sebastiani, 2006) online dictionary which is an extension of WordNet, is referenced. In the SentiWordNet, an emotion is defined as a combination of a positive, objective, and negative rate. For example, the word ‘perfect’ is defined as positive 0.625, objective

22

H.H. Kim et al.

0.25, and negative 0.015. Therefore, the intensity of the tag in the UniEmotion is defined as follows. If a tag is defined as a value greater than 0.75 of positive or negative emotion in the SentiWordNet, the intensity of the tag is strong. If the tag is defined as a value between 0.25 and 0.75 of positive or negative emotion in the SentiWordNet, then the intensity of the tag is middle. Finally, if a tag is defined as a value less than 0.25 in the SentiWordNet, the intensity of the tag is weak. For example, let us take a closer look at tags belonging to the category, happiness. The tag, beautiful, represents a strong intensity of the emotion, while the tag, joyful, represents a weak intensity of the emotion in Figure 2. Also, the tags, favourite and favourites, are considered as a middle intensity of the emotion. The UniEmotion ontology has two classes. UniEmotion:Positive class defines the emotional tags belonging to the four emotional categories of trust, surprise, anticipation, and happiness. UniEmotion:Negative class defines the emotional tags belonging to the four emotional categories of disgust, anger, fear, sadness. The classes have the common properties UniEmotion:Intensity and UniEmotion:Score. UniEmotion:Intensity property specifies the intensity of the tag, and the UniEmotion:Score property specifies the weight of the tag. The values 2.5, 2, 1.5 are assigned for the positive emotional tags according to the intensity of strong, middle, and weak. In the similar way, the values –2.5, –2, –1.5 are assigned for the negative emotional tags according to the intensity of strong, middle, and weak. For the factual tags, the value 1 is assigned.

4.2 Generation of user profiles After the tags are processed using the ontologies, three types of user profiles are generated: listening habits-based profiles, tag-based profiles, and unified profiles. First, the listening habits-based profiles are generated based on the users’ listening habits. Let U1 = {u1, u2, …, um} be a set of users who has listened to music items, and I1 = {i1, i2, …, im} be a set of items, which each user has listened to. Each user ui has a list of items, and these items have preference values. Therefore, a listening habits-based user profile is defined as a ternary relation 〈u, i, p〉, where u ∈ U, i ∈ I1, and p represents the user u’s preference value about item i. In the listening habits-based user profiles, a preference value is calculated by the number of plays of item i. Second, the tag-based profiles are generated based on tags which users have added. Let U2 = {u1, u2, …, um} be a set of users who has added tags to music items, and I2 = {i1, i2, …, im} be a set of items which each user has added tags to. Each user ui has a list of items that he or she has added tags to, and the items have preference values. Therefore, a tag-based user profile is defined as a ternary relation 〈u, i, s〉, where u ∈ U2, i ∈ I2, and s is the weighted value of tags assigned by the UniEmotion ontology. Third, the hybrid profiles are generated by combining the number of plays and score of tags. Let U3 = {u1, u2, …, um} be a set of users who has listened to and has added tags to the music items, and I3 = {i1, i2, …, im} be a set of items which the users have listened to and have added tags to. That is, I3 is defined as I1 ∩ I2. A unified profile is defined as a ternary relation 〈u, i, m〉, where u ∈ U3, i ∈ i3, and preference value m is defined as m = α × p + (1 – α) × s. The value α is chosen 0.5, in which the performance is the best between 0.3 and 0.7.

A unified music recommender system using listening habits

23

4.3 Music recommendation algorithm For the music recommendation, the collaborative filtering algorithm is applied to the user profiles. The first step of collaborative filtering recommendation is finding similar users of a target user. To find similar users of a target user, the Pearson correlation-based similarity is used. The distance between user X and user Y is calculated by (1). r=

∑ ( Xi − X )(Yi − Y ) ∑ ( Xi − X ) ( ∑ (Yi − Y )) 2

2

(1)

Similar users are chosen according to (1), and items that the similar users have rated highly are recommended. The algorithm for the music recommendation is shown in Algorithm 1. Algorithm 1 Music recommendation algorithm Input: a set of user profiles UP {〈u1, i1, n1〉, 〈u1, i2, n2〉, …, 〈u2, i1, n3〉, 〈u2, i2, n4〉, …, 〈um, in, nm〉} Output: a set of recommended items RI{ri1, ri2, …, rin} 1

For all yi ∈ U Compute a similarity s between x and y.

2

Sort by similarity

3

Select top n neighbours Un ← {uk, …, ul}

4

IC = It – Ix

5

For all ui ∈ Un Compute a similarity t between x and ui ∈ Un For all ij ∈ Ic preference +=t* pref

6

Rank by preference

7

Select top n items

Let x be a target user who will receive the recommended items. The input for Algorithm 1 is a set of user profiles UP, which consist of ternary relations . The output of Algorithm 1 is a set of recommended items RI. The algorithm first calculates the similarity s between the target user x and all other users, and selects the top n neighbours among the users, according to similarity (1–3). In the case of tag-based profiles, items Ic are selected as total items It except items Ix that the target user x has already added tags to. In the case of unified profiles, items Ic are selected as total items It except both items Ix that the target user x has already added tags to, and listened to. Next, items Ic are selected as total items It except items Ix that the target user x has already listened to (4). Now, among a set of the similar users Un, compute the similarity t between the target user and similar users. The similarity t is used for the weight of a preference value. The preference value is calculated by the weighted sum of preference value of all similar users (5). Finally, n items are selected according to the highly ranked preference value (6–7).

24

5

H.H. Kim et al.

Performance evaluation

In this section, we show the experimental results for the performance evaluation, and explain the statistical validation in detail.

5.1 Experimental environment For the performance evaluation, a dataset of 1,000 users, 12,600 tags that thy added, and 18,700 music items that they listened to was randomly collected from last.fm. Music items might be albums, artists, or music, and in this study, the artists are used. Apache Mahout is used for the recommender engine. For the user database, MySQL 5.0 is used, and the listening habits collector and the tag score generator are implemented by PHP. UniTag, UniEmotion, and UniMusic ontologies are developed using the protégé ontology editor and knowledge acquisition system.

5.2 Evaluation model Various evaluation models have been considered in the context of recommender systems. In this study, we evaluated precision, recall, and F-measure using the listening habitsbased user profiles, the tag-based user profiles, and the tag-based user profiles. 70% of the whole data are used for the training data, and the remaining 30% are used for the test data. The performance of the recommender system varies with the number of recommended items and similar users. To analyse the effect of the two variables, performance is evaluated with the increasing number of similar users, 2, 4, 8, 16, 32, and 64, and with an increasing number of recommended items, 5, 10, 15, and 20. The definition of precision, recall, and F-measure is as follows. First, recommended items are classified as two groups: items that users are interested in (true positive, TP), and items that users are not (false positive, FP). Similarly, items that are not recommended are classified as two groups: items that users are interested in (false negative, FN), and items that users are not interested in (true negative, TN). The precision is the ratio of items that the users are interested in, to the whole recommended items. It is calculated as (2). precision P =

TP TP + FP

(2)

The recall is the ratio of items that the recommender system recommends, among the items which the user are interested in. It is calculated as (3). recall R =

TP TP + FN

(3)

Finally, the F-measure is the harmonic average of precision and recall. Therefore, the F-measure is calculated as (4). F-measure =

2∗P∗R P+R

(4)

A unified music recommender system using listening habits

25

5.3 Experimental results First, let us consider the precision of the recommendation. To recognise the effect of the number of similar users, experiments are performed with 2, 4, 8, 16, 32, and 64 users, based on the three user profiles. In this case, precisions measure the average precision, when the numbers of recommended items are 5, 10, 15, and 20. Figure 3 shows a change in average precision with increasing number of similar users. The number of similar users is 2, 4, 8, 16, 32, and 64, and in this case, precisions are measured as the average precision when the numbers of recommended items are 5, 10, 15, and 20. Figure 4 shows a change in average precision with the increasing number of recommended items. The numbers of recommended items are 5, 10, 15, and 20, and in this case, precisions are measured as the average precision when the number of similar users is 2, 4, 8, 16, 32, and 64. In all cases, A represents the listening habits-based approach, B represents the tag-based approach, and C represents the unified approach. Next, Figures 5 and 6 show a change in average recalls with increasing number of similar users and with increasing number of recommended items, respectively. In the same way, Figures 7 and 8 show a change in F-measure with increasing the number of similar users and with increasing the number of recommended items, respectively. In all of the evaluation metrics, the unified approach C outperforms the two other approaches. In general, as the number of similar users increases, the performance decreases. However, there is no significant relation between performance and the number of recommended items. Figure 3

Change in precision with the number of similar users (see online version for colours)

26

H.H. Kim et al.

Figure 4

Change in precision with the number of recommended items (see online version for colours)

Figure 5

Change in recall with the number of similar users (see online version for colours)

A unified music recommender system using listening habits

27

Figure 6

Change in recall with the number of recommended items (see online version for colours)

Figure 7

Change in F-measure with the number of similar users (see online version for colours)

28 Figure 8

H.H. Kim et al. Change in F-measure with the number of recommended items (see online version for colours)

5.4 Statistical validation To verify if the experimental results show significant differences, statistical hypothesis testing is used. Let method 1 be the listening habits-based approach, method 2 be the tag-based approach, and method 3 be the unified approach. Since the three groups have asymmetric distributions, one-way ANOVA testing is performed after log transformation. Using a Tukey multiple comparison test, mean differences among the three methods are validated. Table 1 shows the result of a Tukey multiple comparison test for precision. In the Tukey test, groups with significant difference are represented by different characters. As shown in Table 1, method 3 is superior to method 1 and method 2. The difference between method 1 and method 2 is not significant. Table 1

Tukey multiple comparison test for precision

Method

1

2

3

F

Mean of log (precision)

–3.962B

–4.036B

–2.879A

34.27***

Mean precision (standard deviation)

0.020 (0.006)

0.020 (0.009)

0.068 (0.040)

24

24

24

N Note: ***p < 0.001

Table 2 shows the result of the Tukey multiple comparison test for recall. Method 3 is superior to method 1, and method 1 is superior to method 2. The three methods are

A unified music recommender system using listening habits

29

represented by different characters, and therefore they have significant differences in terms of recall. Table 2

Tukey multiple comparison test for recall

Method

1

2 B

3 C

Mean of log (recall)

–3.285

–4.099

Mean recall (standard deviation)

0.044 (0.023)

0.019 (0.010)

0.093 (0.056)

24

24

24

N

–2.635

F A

26.80***

Note: ***p < 0.001

Table 3 shows the result of the Tukey multiple comparison test for F-measure. Method 3 is superior to method 1, and method 1 is superior to method 2. The three methods are represented by different characters, and therefore they have significant differences in terms of F-measure. Table 3

Tukey multiple comparison test for F-measure

Method

1

2 B

3 C

Mean of log (F-measure)

–3.748

–4.117

Mean F-measure (standard deviation)

0.024 (0.006)

0.018 (0.008)

0.064 (0.034)

24

24

24

N

–2.894

F A

41.31***

Note: ***p < 0.001

6

Conclusions and future work

In this paper, we present a unified music recommender system. To use the semantics of tags, the UniTag ontology is developed. In particular, we emphasise that out of tags, emotional tags plays a central role in recommendation process. The UniEmotion ontology classifies the emotional tags into positive emotional and negative emotional tags, and assigns weight according to the intensity. To find the effect of the emotional tags, three types of user profiles are generated. First, the conventional approach is using users’ listening habits. That is, the number of plays of the music items is used for the user profiles. Next, the tag-based user profiles are generated, using the score of each tag. Finally, the unified user profiles are generated, combining both the number of plays, and the score of tags. The music recommendation algorithm is executed based on the three user profiles, and precision, recall, and F-measure are measured. To validate the experimental results, a one-way ANOVA test is performed on precision, recall, and F-measure. Our experiments show that the hybrid approach outperforms the other two approaches with significant differences. Our proposed approach gives a good solution to the conventional cold start problem. Collecting users’ listening habits takes time for users to listen to the music items for a long duration. However, tags are easily added by the users without listening to the music items. For the same reason, the data sparsity problem can be resolved by our approach.

30

H.H. Kim et al.

Tags play a role like word-of-mouth. That is, without directly listening to the music items, the users also add tags to them. Currently, we are elaborating the UniEmotion ontology. In general, context plays a central role in representing the intensity of emotion. However, since tags are used without context, only the tag itself is considered in definition of the emotional tags. Therefore, we have some difficulties in classifying several tags into 8 categories in the UniEmotion ontology. Also, emerging Internet slangs should be handled with care, because Internet slangs may represent positive emotion with negative words. Therefore, research on Internet slangs needs further study.

References Baldoni, M. et al. (2012) ‘From tags to emotions: ontology-driven sentiment analysis in the social semantic web’, Intelligenza Artificiale, Vol. 6, No. 1, pp.41–54. Cantador, I., Konstas, I. and Joemon, M.J. (2011) ‘Categorising social tags to improve folksonomybased recommendations’, Journal of Web Semantics, Vol. 7, No. 1, pp.1–15. Durao, F. and Dolog, P. (2010) ‘Extending a hybrid tag-based recommender system with personalization’, in Proceedings of ACM Symposium on Applied Computing (SAC 2010), Sierre, Switzerland, pp.1723–1727. Esuli, A. and Sebastiani, F. (2006) ‘SentiWordNet: a publicly available lexical resource for opinion mining’, in Proceedings of the 5th Conference on Language Resources and Evaluation, pp.417–422. Firan, C.S., Nejdl, W. and Paiu, R. (2007) ‘The benefit of using tag-based profiles’, LA-Web, Santiago de Chile, pp.32–41. Gruber, T.R. (1995) ‘Toward principles for the design of ontologies used for knowledge sharing’, International Journal of Human Computer, Vol. 43, No. 5, pp.907–928. Gruber, T.R. (2007) ‘Ontology of folksonomy: a mash-up of apples and oranges’, International Journal on Semantic Web & Information Systems, Vol. 3, No. 2, pp.1–11. Kim, H.H. (2012) ‘A tag-based music recommendation using unitag ontology’, Journal of the Korea Society of Computer and Information, Vol. 17, No. 11, pp.133–140. Kim, H.L. et al. (2008), ‘The state of the art in tag ontologies: a semantic model for tagging and folksonomies’, in Proceedings of the International Conference on Dublin Core and Metadata Applications, pp.128–137. Lamere, P. and Pampalk, E. (2008) ‘Social tagging and music information retrieval’, Journal of New Music Research, Vol. 37, No. 2, pp.101–114. Nanopoulos, P., Rafailidis, D., Symeonidis, P. and Manolopoulos, Y. (2010) ‘MusicBox: personalized music recommendation based on cubic analysis of social tags’, IEEE Trans. On Audio, Speech and Language Processing, Vol. 18, No. 2, pp.1–7. Plutchik, R. (1997) ‘The nature of emotions’, American Scientist, Vol. 89, pp.344–350. Raimond, Y., Abdallah, S., Sandler, M. and Giasson, F. (2007) ‘The music ontology’, in Proceedings of the International Conference of Music Information and Retrieval, pp.417–422. Suchanek, F.M., Kasneci, G. and Weikum, G. (2008) ‘YAGO: a large ontology from Wikipedia and WordNet’, Journal of Web Semantics, Vol. 6, No. 3, pp.203–217.

Suggest Documents