In this paper, we describe an approach for creating music recommendations based on user-supplied tags that are augmented with a hierarchical structure extracted for top level genres from Dbpedia. In this structure, each genre is represented by its stylistic origins, typical instruments, derivative forms, sub genres and fusion genres. We use this wellorganized structure in dimensionality reduction in user and item pro ling. We compare two recommenders; one using our method and the other using Latent Semantic Analysis (LSA) in dimensionality reduction. The recommender using our approach outperforms the other. In addition to di erent dimensionality reduction methods, we evaluate the recommenders with di erent user pro ling methods.
These days, most social-networking sites let their members participate in content generation. For example, users can label artists, albums and tracks with tags in Last.fm. A tag can be anything but it is actually a short description of the item. Because tags represent the reason why a listener likes an item, but not how much he/she likes it they are better identi ers of user pro les than ratings, which are usually numerical values assigned to items by users. Thus, WOMRAD 2011 2nd Workshop on Music Recommendation and Discovery, colocated with ACM RecSys 2011 (Chicago, US) Copyright c . This is an open-access article distributed under the terms of the Creative Commons Attribution License 3.0 Unported, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. we concentrate on the tag-based contextual representations of music tracks.
Items are mostly represented in vector spaces in the recommendation systems. In tag-based recommendation systems, users and items are de ned in terms of weighted vectors of social tags. When there is a large amount of tags, calculation of the items to be recommended becomes hard, because working with huge is to represent individual tracks (songs) in lower dimensional spaces. In order to reduce the dimensionality, we focus on the genre information of the tags. Each genre has a relationship with some instrumentation, with some subgenre information and with style information each of which may be entered as tags in the music domain. In our work, for each genre Dbpedia1 (a structured form of Wikipedia2) is crawled to set the relationships between genre and its stylistic origins, typical instruments, derivative forms, sub genres and fusion genres. The contributions of our approach are that: (1) we provide a "semantic relations" method for dimensionality reduction in very huge vector spaces and (2) we perform the comparison of our method against the classical Singular Value Decomposition (SVD) method which is the base of Latent Semantic Analysis (LSA). Our method outperforms the traditional one. 2.
In music recommendation systems, tracks can be pro led
in terms of their audio contents (like rhythm, timbre,
tonality, instrumentation). In addition to audio descriptions,
tracks can be pro led in terms of their text descriptions like
their metadata, lyrics, tags and reviews mined from various
blogs [
Recommender systems either predict ratings for unseen
items or predict items that can be liked. Most of the social
web-sites like Last.fm do not have a rating mechanism.
Instead of explicit ratings, today's recommender systems use
implicit ratings (users' listening habits and purchase
histories etc.). Thus the rating scale in implicit rating
mechanisms is 0-1. Tags can be used in rating-based collaborative
1http://dbpedia.org
2http://en.wikipedia.org
ltering systems with the help of an implicit rating
mechanism [
Up to our knowledge, [
Our system performs 6 main tasks, shown in Figure 1: web crawling, creating an ontology of musical genres, classifying tags according to the ontology, track pro ling, user pro ling and enacting the recommendation process. The circles denote the phases of the system.
Users listen to music and enter tags for tracks in their Last.fm pro les. In the web crawling phase of the system, a data set is generated. Details of the data set are given in Section 4.
Tags may be about genre, instrumentation, location, moods,
style, personal opinions and/or artists' names [
LSA does not use word order and morphology. In order not to di erentiate "electronic" from "electronica", we applied some stemming algorithm. Stemming is a technique to convert similar words to a common base form. This base form does not have to have a meaning from a linguistic point of view (such as reducing synonyms to a single word, or nding the root of the word). Various stemming algorithms exist for the English language. We used the Porter stemmer3 which is a classical stemming algorithm. By using a stemming algorithm, morphology is taken into consideration in our approach. In the tag classi cation phase of our system, we parsed instances existing in the ontology into single words. We applied the stemming algorithm to each single word. Then we concatenated the stemmed roots with "%" in order to consider "word order" that LSA does not use. Some 3http://tartarus.org/ martin/PorterStemmer examples of the stemming results can be seen in Table 3.
All the tags in our dataset are saved in the "tags" table. The reason why we use "%" in the new version of instances is that we use these versions of the instances in our SQL statements. We use the newer instances in SQL statements like "select * from tags where tag name like '%Aborigin%rock%' ". With this usage, we are using about 100000 tags out of 160000 tags in the track representation. In the track proling phase, the size of a track vector is the size of mainstream genres (22 in our case). Last.fm provides integer percentages (between 0 and 100) relative to the most used tags per track. We updated these percentages by adding 1 to each percentage value in order not to discard any having 0 percentage. Each entry in the vector is calculated as follows:
X hasSubGenres(i; j) + X hasF usionGenres(i; j)
Where, g(i; j) is the ith term (genre) in jth track; and hasInstrumentation(i; j) is the total percentage (between 1 and 101) of the tags of the jth track which are found to be similar to the new instance versions of the instrumentation class of ith genre (with the help of the aforementioned SQL statements). The term count is usually normalized to prevent a bias towards longer documents (which may have a higher term count regardless of the actual importance of that term in the document). The term frequency (TF) value gives local information about a tag. An inverse document frequency (IDF) value is calculated for each di erent tag in the training set. This is calculated by dividing the total number of tracks by the number of tracks that refer to that feature. The IDF value gives global information of a tag. Thus, tracks in our dataset are represented as a weighted list of genres and the weights of the genres are calculated with TF*IDF.
wi = ni;t nt log( jT j ) jTij
In the formula above, wi is the weight of ith genre; ni;t equals the number of times ith genre appears in tth track; nt is the total number of genres in tth track; jT j is the total size of the tracks, and jTij equals the number of tracks in which ith genre appears. In the user pro ling phase, 3 di erent methods are used:
1- using the users' own tags (personal tags) that they entered
2- using the users' friends' tags (friends' tags) that their friends entered
3- using all the tags of the tracks (social tags) that they listened to
In the rst method, users are pro led with their own tags. In the second method, users are pro led with their friends' tags. And in the last method, users are pro led with all the tags of the tracks that they listened to. Semantic relations are also used in user pro ling method 1 and method 2, just the same as in track pro ling. In method 3, a user pro le is the sum of the tracks that he/she has listened to. Weights of the genres in user vectors are also calculated with TF*IDF method. The main goal after creating a user pro le from the training set is to recommend the items in the test set.
In the recommendation phase, we use the common cosine similarity method. The cosine similarity formula is given below:
CosSim(track; user) = track vector user vector jtrack vectorjjuser vectorj 4.
We use real Last.fm data in this study. In order to not to use similar users from our own friend lists and in order to achieve diversity, we selected 69 users from an application named "join Last.fm"4. In this group, members of the group share their Last.fm nicknames. We crawled their Last.fm pro les with the help of Last.fm API5. Since our approach is not collaborative but content-based, this number of users is reasonable. Firstly we gathered their top 300 tracks. Then we extracted their "loved" tracks. For each track, we extracted the singer names and tags. We also gathered the 4http://www.facebook.com/group.php?gid=2246697136&v=wall 5http://last.fm/api tag counts per track. Finally for each user we extracted their tags and their friends' tags. Details of our data set can be seen in Table 4.
We performed a 4-fold-cross validation in which the training data size was 75% and the test data was 25%. User pro les were created using the training set and the task of our recommender system was to predict the correct items in the test set. 5.2
In this study we used the most common evaluation metric: Precision at the top N ranked results (P@N). Precision is the ratio of relevant tracks selected correctly to the number of tracks selected. 5.3
In Last.fm, although users listen to music, they rarely
enter tags for the tracks that they like. Thus, user pro les in
Last.fm are smaller than in other social tagging sites, so that
the performance of the pure content-based recommendation
is not satisfying [
User annotated texts, tags in our case, are huge in size, but the representation matrix is very sparse. Using such giant matrices in calculations is a time- and resource- consuming job. For the document categorization or text summarization, LSA has been used for years because it is easy to use and reliable. As an alternative, with the help of Dbpedia, we created an ontology-like semantic relations structure for the music domain. In this paper, we evaluated two methods which can be used in dimensionality reduction. In the evaluation Last.fm dataset was used and the recommenders were evaluated with di erent user pro ling methods. Our method has the advantage of using "word order" and "morphology" with respect to LSA. We plan to extend our work, assigning di erent weights for di erent relations. For instance, hasInstrumentation and hasSubgenres may have di erent weights in the track pro ling.