WAVELET-BASED MUSIC RECOMMENDATION
Claudio Biancalana, Fabio Gasparetti, Alessandro Micarelli, Alfonso Miola and Giuseppe Sansonetti
Department of Computer Science and Automation, Roma Tre University, Via della Vasca Navale 79, Rome, Italy
Keywords:
Wavelet, Music Recommendation, Signal Processing.
Abstract:
Recommender Systems provide suggestions for items (e.g., movies or songs) to be of use to a user. They must
take into account information to deliver more useful (perceived) recommendations. Current music recom-
mender takes an initial input of a song and plays music with similar characteristics, or music that other users
have listened to along with the input song. Listening behaviors in terms of temporal information associated to
ratings or playbacks are usually ignored. We propose a recommender that predicts the most rated songs that
a given user is likely to play in the future analyzing and comparing user listening habits by means of signal
processing techniques.
1 INTRODUCTION
Recommender systems provide suggestions based on
user preferences in order to recommend items likely
to be of interest to a user. It is obvious that user pref-
erences are influenced by the current context, such as
the currenttime of the day, mood, or current activities.
Nevertheless, a few recommender systems explicitly
include this information in the preference models.
A special group of recommender systems are the
ones based on the collaborative approach (Resnick
et al., 1994; Shardanand and Maes, 1995; Breese
et al., 1998). The system generates recommendations
using only information about rating profiles for dif-
ferent users. Collaborative systems locate peer users
with a rating history similar to the current user and
generate recommendations using this neighborhood.
Collaborative filtering (CF) systems have been
successful in several recommender systems. The
availability of large datasets and additional informa-
tion that is easy collectable from the web, makes this
task interesting.
There are several issues that do not allow us to
directly apply the traditional CF approach for music
recommendation. The space of possible items (i.e.,
tracks) can be very large and, similarly, the user space
can also be enormous. Often user ratings are not
available or they cover only a small subset of the user
library of songs. Moreover, when new users enter to
the system or new songs are added to the global li-
brary, it is not possible to provide any recommenda-
tion to them due to the lack of any preference infor-
mation (the so known cold-start problem). There is no
chance to use taxonomies or ontologies to represent
the new items and facilitate the clustering as happens
in different domains (e.g., (Acampora et al., 2010a;
Micarelli et al., 2009)) Content-based approaches col-
lect information describing the items and then, based
on the user preferences, they predict which tracks the
user might enjoy (see for example the Pandora ser-
vice
1
). The key component of this approach is the
similarity function among the songs. Nevertheless,
there is a strong limitation of the highlevel descriptors
that can be automatically extracted from the tracks
(Celma, 2010).
One more relevant issue that traditional CF ap-
proachesdo not take into considerationis the listening
behavior of the user in terms of temporal information.
The timestamp of an item (i.e., when the song song
is played) is an important factor for the recommenda-
tion algorithm. Usually, the prediction function treats
the older items as less relevant than the new ones, but
any further reasoning about the temporal information
is simply ignored.
In this paper, we discuss a recommendation ap-
proach based on signal processing. In particular, a
traditional CF approach is enhanced considering an
improved similarity function between users. The user
listening habits are represented by signals. Wavelet
theory is used to study the related time-frequency rep-
resentations of signals and draw similarity between
listening behaviors. Signal processing techniques are
1
www.pandora.com
399
Gasparetti F., Biancalana C., Micarelli A., Miola A. and Sansonetti G..
WAVELET-BASED MUSIC RECOMMENDATION.
DOI: 10.5220/0003940903990402
In Proceedings of the 8th International Conference on Web Information Systems and Technologies (WEBIST-2012), pages 399-402
ISBN: 978-989-8565-08-2
Copyright
c
2012 SCITEPRESS (Science and Technology Publications, Lda.)
not employed to extract features from the songs, but
for representing and comparing those behaviors in or-
der to group similar users together. This is the novelty
of the approach in comparison with the current litera-
ture.
The rest of this paper is organized as follows. Sec-
tion 2 briefly introduces some related studies on mu-
sic recommendation. Section 3 details our proposed
approaches. Last, in Section 4 a brief account of the
testbed we are developing for the evaluation is given.
Conclusions close the discussion.
2 RELATED WORK
Many algorithms have been developed to address
the personalized recommendation problem. Content-
based approaches aim at including different sources
of information (Semeraro et al., 2009; Groh and
Ehmig, 2007; Micarelli et al., 2006) or better mod-
elling the user interests (Gasparetti and Micarelli,
2007). User-based collaborative filtering (CF) is
widely used, and the main idea is to find the items
liked by other people with similar taste. Different
from the user-based CF, the item-based CF recom-
mends the items which are similar with the user’s
collected items (Schafer et al., 2007). Context-aware
high-level frameworks (e.g., (Acampora et al., 2010b;
Gaeta et al., 2009)) are not easily adaptable to this
specific domain because of the peculiar characteris-
tic of the items. For example, in (Biancalana et al.,
2011a) the authors devise a neural network context-
awarerecommender extracting differentfeatures from
point of interests. In the music scenario, techniques
that automatically extract features from the played
songs are not easily conceivable.
As for music recommendation, the most compres-
sive survey on the literature is to be found in (Celma,
2010). The author groups the recommendation ap-
proachesin four categories: (1) collaborativefiltering,
based on explicit or implicit feedbacks; (2) content-
based filtering, by means of manual or automatic fea-
ture extraction; (3) context-based filtering, the take
advantage of potential user tags associated to each
single song; and (4) hybrid approaches that combine
more then one of the above-mentioned ones.
To the best of our knowledge, there are currently
no attempts to include temporal behavior in user
habits in the music recommendation task. A prelim-
inary attempt has been suggest for the movie domain
in (Biancalana et al., 2011b). The proposed approach
can be categorized as context-based, where the simi-
larity of different songs is evaluated according to the
implicit listening behavior that the user exhibits.
3 WAVELET-BASED
RECOMMENDATION
Traditional user-based CF approaches relies on sim-
ilar users which have similar rating patterns, that is,
the prediction of a rating r
u,s
by user u for the track
track
k
is evaluated as an aggregate of the rating of
some other users for the same item track
k
. We call
these similar users neighbors. If a user v is similar to
a user u, we say that v is a neighbor of u. User-based
algorithms generate a prediction for a track track
k
by
analyzing ratings for track
k
from users in u’s neigh-
borhood.
In order to draw the distance (or similarity) be-
tween two users, the Pearson correlation coefficient is
usually employed (Resnick et al., 1994):
sim(u, v) =
∑
s∈S
u,v
(r
u,s
− r
u
)(r
v,s
− r
v
)
q
∑
s∈S
u,v
(r
u,s
− r
u
)
2
∑
s∈S
u,v
(r
v,s
− r
v
)
2
(1)
where S
u,v
denotes the set of co-rated items between
u and v, r
u,s
is the rating of the user u for the item s,
and r
u
is the average of the ratings of the user u.
Pearson correlation ranges from 1.0 for users with
perfect agreement to -1.0 for users with perfect dis-
agreement. In this way, it is possible to generate a
prediction of rating for the user u and the item s as
follows:
pred(u, s) = r
u
+
∑
v∈NNu
sim(u, v)(r
v,s
− r
v
)
∑
v∈NNu
sim(u, v)
(2)
where NNu is the set of users in the u’s neighborhood.
The proposed recommendation approach is en-
hanced considering a user similarity function that
analyzes contextual factors that are included in the
data collected during the normal usage of the recom-
mender system. In particular, the timestamp associ-
ated to playbacks.
In our recommender we employ Discrete wavelet
transforms (DWT). The basic principles of wavelet
theory were put forth in a paper by Gabor in 1945
(Gabor, 1946). In comparison with the Fourier trans-
form, wavelets are localized in both time (or location)
and frequency instead of just frequency. A wavelet
is a function used to represent a time signal into dif-
ferent scale components. Usually one can assign a
frequency range to each scale component. Each scale
component can then be studied with a resolution that
matches its scale. The DWT is computed by suc-
cessive lowpass and highpass filtering of the discrete
time-domain signal as shown in Fig. 1. This is called
the Mallat pyramid algorithm, a computationally effi-
cient method of implementing the wavelet transform.
WEBIST2012-8thInternationalConferenceonWebInformationSystemsandTechnologies
400
Figure 1: At each decomposition level, the half band filters
produce signals spanning only half the frequency band.
The input signal is assumed to be a set of discrete-
time samples, i.e., a sequence x[n], where n is an inte-
ger. Whereas the basis function of the Fourier trans-
form is a sinusoid, the wavelet basis is a set of func-
tions. In our approach we decide to employ the popu-
lar Haar wavelets.
For an input represented by a list of 2n numbers,
the Haar wavelet transform may be considered to sim-
ply pair up input values, storing the difference and
passing the sum. This process is repeated recursively,
pairing up the sums to provide the next scale: finally
resulting in 2n− 1 differences and one final sum. In
essence, the obtained decomposition can be thought
of as representing a frequency decomposition of the
input.
The algorithm to compute the similarity between
two users is based on the comparison of the Wavelet
transforms obtained by the two signals related to the
listening behavior of the users. In particular, the in-
put dataset is composed of users, songs and tuples
< u
i
,track
k
,timestamp > that represent tracks of a
library L listened by users in a given moment. The
signal is built in the following way:
Algorithm 1: Similarity between users u and v.
for all track
k
∈ L do
vv
u,k
[t
h
] ← number of times the song track
k
has been
played by the user u in the time interval t
h
< t < t
h
+
δT
vv
u,k
[t
h
] ← number of times the song track
k
has been
played by the user v in the time interval t
h
< t < t
h
+δT
end for
for all track
k
∈ L do
w
u,k
← discrete Haar Wavelet transform of signal vv
u,k
w
v,k
← discrete Haar Wavelet transform of signal vv
v,k
end for
sim
u,v
← Euclidean distance between the two vectors w
u
and w
v
Two users are considered similar if they listen to
the same songs in the same time of the day. If the two
users listen to the same songs in a given period of time
but this two periods do not coincide, e.g., the user u
played some songs in January and v played the same
songs in March, traditional comparison metrics return
low similarity between the two users.
The Euclidean distance between the coefficients
of the two wavelets allows us to ignore potential time
shifts between listening behaviors. Moreover, it takes
into account the frequency of items, i.e., the times
a song has been played in a given period. In this
way, we are able to recognize similarities between
user habits analyzing different scales or approxima-
tions of the input signals produced by the wavelet tree
decomposition. This is a well-known characteristic of
the wavelet theory that is exploit several times in the
Content-based image retrieval domain.
4 HOW TO EVALUATE
We are currently devising a testbed that include
enough real usage data extracted by a public domain
datasets, i.e., Last.fm Dataset - 1k
2
, in order to com-
pare the performance with traditional recommender
approaches by means of standard evaluation mea-
sures. That dataset contains tuples in the following
form:
< user, timestamp, artist, song > (3)
collected from Last.fm website that correspond to
one song played by a specific user. There are 992
unique users and more than 19 Million entries, there-
fore enough data to represent user listening habits.
The pair < artist, song > is easily mapped to the track
variable used in the wavelet-based recommender. We
will looking to standard measures of prediction be-
tween the recommended songs and the actual songs
played by the users.
5 CONCLUSIONS
In this paper, we have discussed a recommender ap-
proach based on signals related to the user listening
behavior.
Further work will be done along several research
directions. Some factors that should be included in
the recommendation process are the novelty of songs
and the user authority. New songs have a higher po-
tential of being interesting than old songs. Moreover,
some collaborative approaches have tried to diversify
the ratings from users, identifying more authoritative
2
last.fm
WAVELET-BASEDMUSICRECOMMENDATION
401
users that should be taken more into consideration
when predictions have to be suggested. Serendipity
and negative preferences are further factors that mu-
sic recommenders should include in their predictive
analysis (Iaquinta et al., 2008; Musto et al., 2011).
REFERENCES
Acampora, G., Gaeta, M., and Loia, V. (2010a). Exploring
e-learning knowledge through ontological memetic
agents. Comp. Intell. Mag., 5:66–77.
Acampora, G., Gaeta, M., Loia, V., and Vasilakos, A. V.
(2010b). Interoperable and adaptive fuzzy services for
ambient intelligence applications. ACM Trans. Auton.
Adapt. Syst., 5:8:1–8:26.
Biancalana, C., Flamini, A., Gasparetti, F., Micarelli, A.,
Millevolte, S., and Sansonetti, G. (2011a). Enhanc-
ing traditional local search recommendations with
context-awareness. In Konstan, J. A., Conejo, R.,
Marzo, J. L., and Oliver, N., editors, User Model-
ing, Adaption and Personalization - 19th International
Conference, UMAP 2011, Girona, Spain, July 11-15,
2011. Proceedings, volume 6787 of Lecture Notes in
Computer Science, pages 335–340. Springer.
Biancalana, C., Gasparetti, F., Micarelli, A., Miola, A., and
Sansonetti, G. (2011b). Context-aware movie recom-
mendation based on signal processing and machine
learning. In Proceedings of the 2nd Challenge on
Context-Aware Movie Recommendation, CAMRa ’11,
pages 5–10, New York, NY, USA. ACM.
Breese, J. S., Heckerman, D., and Kadie, C. M. (1998). Em-
pirical analysis of predictive algorithms for collabora-
tive filtering. In Cooper, G. F. and Moral, S., editors,
Proceedings of the 14
th
Conference on Uncertainty in
Artificial Intelligence, pages 43–52.
Celma, . (2010). Music Recommendation and Discovery -
The Long Tail, Long Fail, and Long Play in the Digital
Music Space. Springer.
Gabor, D. (1946). Theory of communication. J. Inst. Elect.
Eng., 93:429–457.
Gaeta, A., Gaeta, M., and Ritrovato, P. (2009). A grid based
software architecture for delivery of adaptive and per-
sonalised learning experiences. Personal Ubiquitous
Comput., 13:207–217.
Gasparetti, F. and Micarelli, A. (2007). Personalized search
based on a memory retrieval theory. International
Journal of Pattern Recognition and Artificial Intel-
ligence (IJPRAI): Special Issue on Personalization
Techniques for Recommender Systems and Intelligent
User Interfaces, 21(2):207–224.
Groh, G. and Ehmig, C. (2007). Recommendations in taste
related domains: collaborative filtering vs. social fil-
tering. In GROUP ’07: Proceedings of the 2007 inter-
national ACM conference on Supporting group work,
pages 127–136, New York, NY, USA. ACM.
Iaquinta, L., de Gemmis, M., Lops, P., Semeraro, G., Fi-
lannino, M., and Molino, P. (2008). Introducing
serendipity in a content-based recommender system.
In Xhafa, F., Herrera, F., Abraham, A., K¨oppen, M.,
and Ben´ıtez, J. M., editors, International Conference
on Hybrid Intelligent Systems (HIS 2008, pages 168–
173. IEEE Computer Society.
Micarelli, A., Gasparetti, F., and Biancalana, C. (2006).
Intelligent search on the internet. In Stock, O. and
Schaerf, M., editors, Reasoning, Action and Inter-
action in AI Theories and Systems, volume 4155 of
Lecture Notes in Computer Science, pages 247–264.
Springer.
Micarelli, A., Sciarrone, F., and Gasparetti, F. (2009). A
case-based approach to adaptive hypermedia naviga-
tion. IJWLTT, 4(1):35–53.
Musto, C., Semeraro, G., Lops, P., and de Gemmis, M.
(2011). Random indexing and negative user prefer-
ences for enhancing content-based recommender sys-
tems. In Huemer, C. and Setzer, T., editors, E-
Commerce and Web Technologies, volume 85 of Lec-
ture Notes in Business Information Processing, pages
270–281. Springer.
Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P., and
Riedl, J. (1994). Grouplens: an open architecture
for collaborative filtering of netnews. In Proceedings
of the 1994 ACM conference on Computer supported
cooperative work, CSCW ’94, pages 175–186, New
York, NY, USA. ACM.
Schafer, J. B., Frankowski, D., Herlocker, J., and Sen, S.
(2007). Collaborative filtering recommender systems.
In Brusilovsky, P., Kobsa, A., and Nejdl, W., edi-
tors, The Adaptive Web: Methods and Strategies of
Web Personalization, volume 4321 of Lecture Notes
in Computer Science, pages 291–324. Springer Berlin,
Heidelberg, Berlin, Heidelberg, and New York.
Semeraro, G., Lops, P., Basile, P., and de Gemmis, M.
(2009). Knowledge infusion into content-based rec-
ommender systems. In Bergman, L. D., Tuzhilin, A.,
Burke, R. D., Felfernig, A., and Schmidt-Thieme, L.,
editors, Proceedings of the 2009 ACM Conference on
Recommender Systems, RecSys 2009, pages 301–304.
ACM.
Shardanand, U. and Maes, P. (1995). Social information fil-
tering: algorithms for automating word of mouth. In
Proceedings of the SIGCHI conference on Human fac-
tors in computing systems, CHI ’95, pages 210–217,
New York, NY, USA. ACM Press/Addison-Wesley
Publishing Co.
WEBIST2012-8thInternationalConferenceonWebInformationSystemsandTechnologies
402
