=Paper=
{{Paper
|id=Vol-1245/paper5
|storemode=property
|title=A User-centered Music Recommendation Approach for Daily Activities
|pdfUrl=https://ceur-ws.org/Vol-1245/cbrecsys2014-paper05.pdf
|volume=Vol-1245
|dblpUrl=https://dblp.org/rec/conf/recsys/DiasFC14
}}
==A User-centered Music Recommendation Approach for Daily Activities==
https://ceur-ws.org/Vol-1245/cbrecsys2014-paper05.pdf
A User-centered Music Recommendation Approach for
Daily Activities
Ricardo Dias Ricardo Cunha Manuel J. Fonseca
INESC-ID, Instituto Superior Técnico, Faculdade de Ciências,
Universidade de Lisboa Universidade de Lisboa
Lisboa, Portugal Lisboa, Portugal
{ricardo.dias, mjf@di.fc.ul.pt
ricardo.d.cunha}@tecnico.ulisboa.pt
ABSTRACT 1. INTRODUCTION
The number of songs available on the Internet has grown Over the last decade, due to the increasingly easy access
steadily over the last decade, with the recent growth being to online music streaming services, people have gained the
due mainly to streaming services. As a consequence, it is ex- opportunity to listen to millions of songs over the Internet
tremely difficult for users to find the appropriate music that and almost everywhere. However, the opening to this broad
suit their needs, in particular, while using systems that do spectrum of songs lead people to feel paralyzed and doubtful
not have any previous information about them. This is fur- [17], as it gets harder for them to filter out the songs they
ther exacerbated while selecting appropriate songs for daily would enjoy the most, specially while performing other ac-
activities, like shopping, running or sleeping. In this paper tivities, such as practicing sports, driving or studying [11].
we describe Improvise, a personalized music recommenda- Due to the large number of songs that users have access, it is
tion solution for daily activities, whose approach associates hard for them to select the most appropriate songs for their
music content (acoustic features) with activities (context). activities.
Each activity is characterized by determining intervals for
each content feature, which are then used to filter out songs To reduce the burden of choosing among too many songs,
to be suggested to users. While the initial intervals are researchers have focused on creating recommender systems
generic enough to provide recommendations for di↵erent ac- that can automatically generate recommendations that fit
tivities without having previous knowledge about the user’s users’ preferences. Celma’s extensive work on recommenda-
tastes, our approach also considers users’ feedback to person- tion [5] classifies recommender systems into five typical cate-
alize the recommendations for each user and activity. This gories: Demographic Filtering, Collaborative Filtering, Con-
is done by adapting the intervals according to the feedback text Aware Filtering, Content-Based Filtering and Hybrid
from users. Preliminary evaluation shows that we are on Methods. Recently, hybrid and context-aware approaches
the good path to achieve the goal of developing a solution have gained relevance amongst researchers, as they agree
to e↵ectively recommend songs for daily activities, and able that listening patterns can be influenced by di↵erent fac-
to adjust to individual user’s tastes, increasing their satis- tors, such as temporal properties [7], location, emotions and
faction. the activity a listener is engaged in [21].
Categories and Subject Descriptors Concerning music selection for daily activities, several ap-
H.3.3 [Information Search and Retrieval]: Retrieval proaches have been proposed. In [21], the authors created
Models a mobile system that is able to detect what activity the
user is performing and select the appropriate music for it,
based on the time of the day, accelerometer data, and audio
General Terms from the the microphone. Lifetrak [15] is a context-aware
Design, Algorithms, Human Factors, Experimentation playlist generator that automatically chooses music in real-
time based upon the location, the pace of movement, the cur-
Keywords rent time, and other phenomena in the users environment.
User-centered, Music Recommendation, Content, Context, It uses a simple learning mechanism to adjust the ratings
Daily Activities, Cold-Start of songs for a particular context based on users feedback
when a song is being played. However, these approaches
are still very impersonal with little control, lacking users in-
volvement through the whole steps of the recommendation
process, which could improve the users satisfaction and con-
fidence [9, 20].
In this work we describe Improvise, a user-centered recom-
Copyright 2014 for the individual papers by the paper’s authors. mendation approach for daily activities. Improvise is a rec-
Copying permitted for private and academic purposes. This volume is ommendation model developed by characterizing activities
CBRecSys 2014,
published and October 6, by
copyrighted 2014, Silicon Valley, CA, USA.
its editors. in terms of content features, by defining their boundaries.
Copyright
CBRecSys 2014
2014,byOctober
the author(s)
6, 2014, Silicon Valley, CA, USA.
26
The set of intervals generated for each feature and activ- to characterize the listening process [1], such as, the place
ity are then used to filter out songs to be suggested. We where we are listening to the music, the time of day, the
developed a generic model with this approach (to be used activity we are performing, etc. Context aware recommen-
initially by all users) by gathering data from several users. dation systems (CARS) use context information to describe
Individual personalization of this model is created over time and characterize the songs or artists we listen to. For ex-
by taking into account user’s feedback for each activity, and ample, Su et al. [19] improved Collaborative Filtering (CF)
consequently adapting the intervals based on the features methods combining user grouping by location, motion, cal-
from the new selected songs. endar, environment conditions and health conditions, while
using content analysis to assist the system in the selection
Preliminary experimental evaluation revealed that the ini- of the appropriate songs. On the other hand, Park [14] de-
tial generic model was able to suggest songs to daily activi- veloped a modified User-based CF method (called Session-
ties for di↵erent users, without any knowledge about them. based CF), where users were replaced by sessions, adding a
Moreover, by using the user feedback (selected songs from temporal dimension to CF recommendations. In [12], Liu
the recommendation list) to personalize the recommenda- et al. took the change in the interests of users over time
tion model, increased the number of adequate songs for each into consideration and added time scheduling to the music
activity in 25% (while compared to the generic model) and playlist. Baltrunas et al. [2] introduced a new context-aware
the users satisfaction. recommendation approach called user micro-profiling, where
the user profile is split into several sub-profiles, each one
We highlight as main contributions the following: i) a user- representing the user in a particular context. The authors
centered personalized solution to perform recommendations stated that the choice of songs during the day is influenced
for daily activities; ii) an alternative approach for solving by contextual conditions, such as, the time of day, mood or
the cold-start problem. the current activity listeners perform. In [23], the authors
presented a novel and improved statistical model for charac-
In the next section we discuss the related work, presenting terizing user preferences in consuming social media content.
research made in music recommender systems. Section 3 By taking into account information about listening sessions
describes the proposed solution for music recommendation, of individual users, they have arrived at a new session-based
and in Section 4 we present the experimental user evaluation hierarchical graphical model that enhanced individual user
and discuss the preliminary results of our work. Finally, in experience.
Section 5 we present future work and conclude this research.
In short, there have been some e↵ort from researchers to
2. RELATED WORK create automatic mechanisms that characterize users prefer-
ences through the use of di↵erent sorts of data, like tempo-
Several approaches for music recommendation have been de-
ral patterns, emotions, or choices behind song selection for
veloped so far. Depending on the type of data used to per-
particular activities. On the other hand, content features
form the recommendations, we can categorize the methods
have been extensively used for recommendation and playlist
in di↵erent groups [5], either if they use demographic data
generation because of the benefits they present. Despite
[22], listening habits and ratings [10, 6], content informa-
that, there has been little work on engaging users through
tion from songs [4], the context when they were listened [12,
the recommendation process, giving them the control over
19], or any combination of the previous [18, 16]. Although
how profiles are created and managed. We intend to tackle
collaborative filtering solutions have been the most widely
this gap by developing a recommendation approach based
researched techniques in the past, nowadays, a huge e↵ort
on user input and feedback as well as on content features,
has been put on techniques that capture the context around
for creating a customizable recommendation model for each
the listening activity [16], because they can provide insights
user.
about the reasons that lead users to listen to certain songs.
Content-based approaches use the description of songs to 3. IMPROVISE
compute similarities and recommend songs similar to the In this research we describe Improvise, a personalized rec-
user favorite ones or to chosen seeds. These approaches ommendation system able to suggest songs that fit the users
solve the early-rater and popularity bias problems, as all the needs while performing daily activities. To achieve this goal
items are considered to be of equal importance [5] (without we followed a user-centered approach, taking advantage of
human intervention). However, a potential problem of these users’ input and feedback to develop a generic model capable
approaches is the novelty problem. Assuming that the sim- of recommending songs to everyone, even without any pre-
ilarity function works accurately, one might assume that a vious knowledge about them. Activities were characterized
user will always receive items too similar to the ones in his using content-based features. Five activities were considered
profile. To cope with this problem, recommenders should based on the existing related work [21]: walking, relaxing,
use other factors to promote the diversity and novelty of running, sleeping and shopping.
the recommended items. In the solution proposed by Cano
[4], acoustic features of songs (timbre, meter and rhythm pat- 3.1 Approach Overview
terns) were used for recommendation. Daniel M. [13] used Figure 1 shows the di↵erent steps for creating the recom-
lyric feature analysis to find similar items that describe race mendation models, explaining the recommendation process.
conflicts and social issues. In [3], Cai recommends music First, we associate songs with activities by using the user in-
based only on emotion. put gathered through a web-application (Figure 1-1). This
allowed us to analyze what songs were more suitable for each
Context can be defined as any information that can be used activity (based on the users preferences) and thereby create
26
Figure 1: The di↵erent steps in Improvise to generate recommendations.
the broad and generic recommendation model. To this end, used by the majority of digital music services (like Musicov-
we then gathered content features from the songs, using the ery2 ), showing only 15 di↵erent genres (Rock, Electro, Pop,
EchoNest 1 service (Figure 1-2). The values for each feature R&B, Rap, Metal, Classic, etc.). After selecting one or more
were collected to characterize each activity, by inferring in- genres suitable for each activity, users could choose from 30
tervals of values for each feature. These intervals define a artists in maximum (twice the number of genres). Next,
set of hyper-rectangles (Figure 1-3), which Improvise uses to users could finally choose songs from the artists previously
generate recommendations by filtering out songs from the selected (a maximum of 100 songs were shown). Top artists
EchoNest service (Figure 1-4). and songs were used for the selection, gathered through the
online service EchoNest. The result of this process was an
To personalize the recommendation model we consider the association between activities and a set of songs suitable to
songs selected by the users as appropriate for each activity be listened by di↵erent users.
(user’s feedback), extract the content properties for these
new songs and recalculate the hyper-rectangles for each ac- To collect this data we sent emails to contacts and spread
tivity, repeating steps 2 and 3. By using this approach our the link for the application through social networks, namely,
solution adapts the recommendations to the users’ tastes Facebook and Google+, to reach as many users as possible.
and preferences over time. 98 subjects used the application, providing a total of 251 an-
swers for all the activities. Despite the fact that some users
In the following sections we provide details for the di↵erent did not provide feedback about their tastes for all activities,
steps described here. the distribution was uniform: 55 answers for the activity
walk, 53 for running, 47 for sleeping, 48 for relaxing and 48
for shopping. This resulted in associating 8,518 songs with
3.2 Association between songs and activities the activities.
To develop a music recommendation system able to suggest
songs suitable for di↵erent activities is necessary to capture To characterize the activities we extracted content informa-
the users’ tastes and preferences for those activities. For tion from the songs selected by users using the Echonest
example, to understand the reasons ”why do users select service. For performance issues, we opted to use only the
certain songs for running, and others for relaxing?”. Al- top-100 preferred songs for each activity.
though di↵erent criteria can influence this selection, some
conceptual properties are shared among users for the same After extracting all the features o↵ered by EchoNest, we per-
activities, such as, familiarity or distraction. However, these formed an evaluation to measure how discriminative each
sort of more subjective features are difficult to extract and feature was in this characterization. This evaluation was
encode. On the other hand, content-based features have performed using the CfsSubsetEval attribute evaluator along
been used for some time in retrieval and recommendation with the best-fit search method from Weka3 [8]. The follow-
systems [4, 13], presenting some advantages: they can be ing four features were selected as the most discriminative:
automatically extracted and used to compute the similarity accousticness, energy, loudness and tempo. Therefore, for
between songs; they help solving the problem of cold-start each song we created a 4-feature vector describing its con-
for new songs. Therefore, they constitute a good approach tent, and for each activity an array of feature vectors of the
to characterizing activities and empower a recommendation songs associated with them (see Figure 1-2).
solution.
To associate songs with activities and thus characterize them 3.3 Generic Recommendation Model
by using content-based features, we took a user-centered ap- The association detailed in the previous subsection allows
proach. To that end, we developed a web-application to us to describe each activity through a set of feature-vectors,
collect songs that users enjoy listening to, while perform- representing each vector a song chosen by the users. To
ing each activity (see Figure 1-1). The users selected songs use this information in the suggestion of songs we need to
first by filtering genres, then artists, and finally by songs convert it into a simpler representation to facilitate the rec-
(see Figure 2). Regarding genres, we adopted the taxonomy
2
http://musicovery.com
1 3
http://the.echonest.com/ http://www.cs.waikato.ac.nz/ml/weka/
27
(a) Genre selection.
Figure 3: Method for determining the hyper-rectangle lim-
its.
viding an adaptable recommendation mechanism as detailed
later in the paper. The hyper-rectangles represent the back-
bone of Improvise. To recommend songs using them, we
search for songs within the limits of the hyper-rectangles,
(b) Artist selection. using the EchoNest service.
Our generic recommendation approach consists in creating
five generic hyper-rectangles, one for each activity, based on
the songs collected through the web-application developed
(see Section 3.2). This generic model is therefore capable
of suggesting songs for each activity to any user, without
having previous knowledge about him/her. This way we
have a simple approach that provide an answer to the cold-
start problem.
To calculate the intervals for each feature and thus define the
hyper-rectangles, we started by testing two di↵erent meth-
(c) Songs selection. ods. The first method (M1 ) used the average minus the
standard deviation for finding the minimum of the interval
Figure 2: Web-application developed to associate songs with and the average plus the standard deviation to find its max-
activities. imum. The second method (M2 ) used the 10% percentile as
the minimum value of the interval and the 90% percentile
for its maximum. To evaluate the quality of the two meth-
ommendation process. ods, we searched for songs within the intervals defined, using
the minimum and maximum values of the intervals for each
Typically, a recommendation system is often seen as a sug- feature. The results of these tests lead us to conclude that
gestion of items similar to a feature vector that represents the methods were not adequate since the intervals generated
the users’ tastes. However, this approach is very restrict, were too wide, with a considerable overlap between them,
since it will tend to recommend the same set of songs every blurring the di↵erences between the recommendations for
time. Based on this and on the mechanism we are using the di↵erent activities.
to characterize each activity, we decided to use an interval
approach for the recommendation. Therefore, we created two new methods: the first based on
M1 using a percentage of the standard deviation, with val-
To this end, we defined a set of intervals delimiting the value ues of 15, 20, 25 and 30%; and the other method, similar
that each feature could take, instead of considering a single to the previous one, but using the median instead of the
point in space (computed using a clustering algorithm, for average (M4 ). The limits for the hyper-rectangles were gen-
instance). Moreover, this approach not only increases the erated in two di↵erent ways: one using the top-100 songs se-
range of songs that we can suggest for each activity, but lected by users, and the other using only the top-20. These
also gives the possibility of using di↵erent sub-intervals to variants, generated a set of 8 di↵erent data sets that were
restrict the filtering process. This set of intervals define what used to assess the quality of the proposed methods for the
we labeled as the hyper-rectangle (see Figure 1-3). A hyper- hyper-rectangle calculus. The datasets were used for train-
rectangle has four dimensions, and is defined by intervals ing a Random Forest classifier with the goal of evaluating
with a maximum and a minimum value that each feature the quality of the limits generated (the adequacy of the songs
can take within each activity. The size and position of these to the activity). Figure 3 depicts the accuracy values of the
rectangles di↵er between activities and for each user, pro- classifiers. In this figure, C1 and C2 encode the datasets
28
used to train the Random Forest classifiers: C1 represents web application where users could select the songs they con-
the top-20 songs dataset, and C2 the top-100. Notice that sider appropriate for each activity. By counting the number
C1 is a subset of C2, as these songs were those selected by of suitable songs we could measure the e↵ectiveness of Im-
users. The labels S1 and S2 encode the datasets used for provise in suggesting songs for daily activities.
determining the intervals of the hyper-rectangles: S1 indi-
cates that the top-20 songs were used, while S2 represents In the following sections we describe our objectives, the par-
the usage of the top-100 songs dataset. Again, S1 is a subset ticipants, the evaluation procedure, the main results and the
of S2. Finally, AVG stands for the average, while MEDN for discussion about them.
the median.
We used these two dataset divisions (S1 and S2) to under-
4.1 Goals and Tasks
The main goal of Improvise is to recommend and suggest
stand if it would be beneficial to have a wider (100) or nar-
songs to be listened while doing activities, such as, running,
rower (20) history of songs for generating the intervals. Al-
relaxing or shopping. To validate both the generic and the
though the best result was achieved with the average ±15%
personalized solution, we divided the evaluation into two
of the standard deviation, the number of songs suggested by
phases.
this method was smaller while compared to others. There-
fore, we chose the second best method, which corresponds to
The first phase consisted in evaluating the generic model to
the usage of the median ±20% of the standard deviation. In
understand if it was flexible enough to recommend music
this case the median and standard deviation were calculated
that fit the preferences of any potential user. In the sec-
using the top-100 songs chosen by the users. The training
ond phase, the songs selected by each user during the first
instances used by the classifier were the top 20 songs chosen
evaluation (feedback) were used to individually personalize
by the users for each activity (C1-S2-MEDN ).
Improvise and to generate new recommendations for each
activity. The main objective consisted in understanding if
In summary, to create the generic recommendation model
personalized suggestions were better than those generated
we defined a set of five hyper-rectangles, one for each ac-
using the generic model. Finally, a second interaction with
tivity, using the top-100 songs and the median ± 20% of
the personalized model was conducted to assess the impact
the standard deviation as the method to determine their in-
in personalization over time. Here, the personalized recom-
tervals. Thus, without previous knowledge about a user’s
mendation model su↵ered a second personalization by taking
preferences, we can generate recommendations suitable for
into account the new feedback collected during the previous
him/her and for the activity at hand (see Figure 1-4).
session.
3.4 Personalized Recommendation Model The main task for both phases consisted in selecting the
To personalize the recommendations for each activity we appropriate songs for each activity from a list of songs sug-
incorporate the user feedback, expressed by selecting the gested by our solution.
songs she/he considered adequate for the activity. This is
then materialized by adjusting the intervals for each activity
based on the songs listened. 4.2 Participants
During the first phase of the evaluation, ten users parti-
While the method for determining the hyper-rectangles in cipated in the experiment. Eighty percent of the subjects
the personalized model is the same as in the generic ap- were male, with ages between 22 and 29 years old (90%), be-
proach (top-100 songs and the median ± 20% of the stan- ing graduate or undergraduate students from the university
dard deviation), the list of songs used is di↵erent. This list campus. All of them reported listening to music for di↵erent
starts with the top-100 songs chosen by all users (and used activities during the day.
to create the generic model) and is updated with the new
songs selected by the users. These are added to the end In the first iteration of the second experiment all the ten
of the list replacing the oldest ones, as they represent less previous subjects participated in the tests using their per-
preferred songs. sonalized version of the hyper-rectangles for each activity,
created based on their feedback from the first phase. Due to
When the list of songs used to generate the intervals no time restrictions, only five of the ten users were able to par-
longer contains songs used for the generic model, the process ticipate in the second iteration of the second phase. Here,
follows a FIFO order (First In First Out). This approach we used a new personalized version of the recommendation
constantly personalizes the recommendation model by con- model, created using the feedback provided in the previous
sidering the user feedback and by adjusting to his/her cur- session.
rent tastes and preferences, over time. New songs remain
more time in the list used to determine the new intervals. 4.3 Procedure
This design allows us to perform a more personalized rec- To evaluate the proposed solution we developed a web appli-
ommendation, taking advantage of the current tastes and cation for users to interact with the recommendation tech-
preferences of the users. nique (see Figure 4). For both experiments, the evaluation
started first with users answering a small questionnaire with
4. EVALUATION demographic information to characterize them (e.g. age,
We conducted two user-centric experiments to evaluate both gender, music listing information, etc.). Then users selected
recommendation approaches o↵ered by Improvise, the generic the appropriate songs for each activity, and at the end they
and the personalized model. To that end, we developed a filled a satisfaction questionnaire. Notice that the activities
29
(a) Activity selection. (b) Song selection.
Figure 4: Application developed for evaluating the recom-
mendation model.
Figure 6: Results of the user satisfaction regarding songs
were simulated, as the users were not actually performing suggested by the generic model.
them, we just mentioned their names.
To evaluate the adequacy of the generic and personalized per activity (30%). This corresponds to an average increase
models, we presented 50 songs for each activity, from which of 25% over the number of songs selected using the generic
users should select those they consider correctly assigned to model. Sleeping is the activity that presents the best results
the activity. Songs were presented (album cover, song and and the highest improvement for the personalized model.
artist name) one at a time, with the possibility of playing a
30 seconds sample. The satisfaction questionnaire used to collect users opinion
about the quality of the suggested songs was composed by
After selecting the songs for each activity, users were asked a five point likert scale, with answers as strongly disagree,
to answer a satisfaction questionnaire to express their agree- disagree, neutral, agree and strongly agree. Figure 6 depicts
ment with the suitability of the suggested songs for the ac- the results for the generic model. Overall, more than half of
tivity in question. the users agreed or strongly agreed with the suggested songs,
for four of the five activities. Only the Shopping activity did
not achieve this value.
4.4 Results
Overall, users were satisfied with the recommendations per- Figure 7 depicts the total number of songs considered ap-
formed by both the generic and the personalized model. propriate for the various activities. As we can see, there is
Moreover, the e↵ectiveness of the personalized model was a steady increase in the number of correct songs, from the
confirmed by a steady increase in the number of songs con- generic model to the second iteration of the personalized
sidered suitable for each activity by the users, from the model. Indeed, this corresponds to an increase of 31% (on
generic model to the personalized model. average) for all users, revealing that our model is able to fit
the tastes of the di↵erent users over time.
On average users selected eleven to twelve songs, for each
activity, from the list suggested by the generic model, cor- The growth in the number of songs from the first to the
responding to 24% of the total of songs recommended. second iteration of the personalized model was around 10%.
Detailed data on the behavior of the three models, for the
For the first iteration of the personalized model, as depicted five users who participated in the three test sessions, is de-
in Figure 5, users selected on average more than 15 songs picted in Figure 8. As we can see, overall, the personalized
Figure 5: Comparison between the generic and the person- Figure 7: Evolution of the total number of songs selected for
alized model in terms of the number of songs selected for the various activities using the di↵erent recommendations
each activity. Error bars denote standard deviation. models.
30
Figure 8: Results of the evaluation of the personalized model Figure 9: Results of the user satisfaction regarding songs
for each user. suggested by the personalized model during the first itera-
tion.
models suggested more songs suitable for the di↵erent activ-
In a particular case the personalized model required two it-
ities than the generic model (Users 1, 2, 4 and 5). Only for
erations to get adjusted to the user, showing that for some
User 3 the personalized model suggested less adequate songs.
users our model needs more time to ”learn” the users pref-
Two other special cases are worth mentioning: for User 2
erences. In another case, for which we did not find any evi-
the second iteration with the personalized model performed
dence for it, the user preferred more songs from the generic
worse than in the first iteration; and for User 5, the person-
model than from the personalized ones.
alized model required a second iteration to outperform the
generic model.
Although these results are very promising, showing that our
approach can deal with the cold-start problem by providing
Figure 9 depicts the satisfaction results for the first iteration
a generic model that can suggest songs for any user without
of the personalized model. Similarly to what happened with
knowing anything about them, we would like to mention
the generic model, more than half of the users agreed or
some constraints that prevent us from state stronger claims.
strongly agreed with the suggested songs, for four of the five
First, we cannot draw any statistical significance from the
activities. But, for the personalized model we have more
results due to the small number of users involved in the
strongly agree answers. The Shopping activity still has the
preliminary evaluation. In a near future we plan to perform
worst results, but are better than in the generic model.
an evaluation with a larger number of users. Second, users
were not performing the activities for which we suggested
To get a better understanding of the improvement provided
songs. More evaluation is required to clarify if this a↵ected
by the personalized model, we grouped the users’ answers
the result.
about the generic and the personalized model in negative
(strongly disagree and disagree), neutral and positive (agree
and strongly agree) opinions. We found an increase of 13% 5. CONCLUSIONS AND FUTURE WORK
in the number of positive opinions from the generic to the Nowadays, a huge amount of songs is available to millions of
personalized model, showing that the personalized sugges- users around the world. With millions of artists and songs
tions are more inline with users’ preferences. on the market, it is difficult for users to find songs that
please them. This problem is even worse when trying to
select songs for di↵erent activities.
4.5 Discussion
From these preliminary results, we can conclude that our In this paper we described Improvise, an adaptable solution
work is on the good path to create an approach able to e↵ec- for recommending songs for daily activities. Improvise is
tively suggest songs for daily activities and flexible enough to a user-centered approach that relies on the hyper-rectangle
adapt the recommendation list to the users’ tastes and pref- concept, determined using content from songs. We described
erences over time, supporting both ”unknown” and ”known” the rationale behind the calculus of the hyper-rectangles
users. for a generic recommendation model and also the creation
of a personalized solution. Preliminary results show that
Results for the number of songs chosen for each activity the generic model was successful in recommending songs to
show that users selected more songs while using the per- users. But more relevant is the flexibility of the solution in
sonalized model than while using the generic model. This adapting the recommendation to di↵erent users for each ac-
confirms that our solution can e↵ectively suggest songs for tivity, increasing not only the number of songs selected, but
di↵erent users and activities, and adapt to their preferences. also their satisfaction.
Moreover, the second iteration with the personalized model
reinforced these results. Satisfaction results were also in Regarding future work, we plan to explore two paths. The
agreement with the reported increase in the number of songs first is to explore new and di↵erent methods for determining
selected. Users were overall happy and satisfied with the rec- the hyper-rectangles, like for instance to consider more than
ommendations performed. one hyper-rectangle for each activity. This can capture more
31
diverse and sparse preferences and tastes, promoting new Conference on Recommender Systems, 2012.
recommendations and user satisfaction. The second path [10] J. A. Konstan, B. N. Miller, D. Maltz, J. L. Herlocker,
is to use a larger number of songs to determine the hyper- L. R. Gordon, J. Riedl, and H. Volume. Grouplens:
rectangles that characterize the user profile, since at the Applying collaborative filtering to usenet news.
moment we are only using the top-100 songs preferred by Communications of the ACM, 1997.
the users as detailed in Section 3. Although, using more [11] T. Leong, S. Howard, and V. Frank. Choice:
songs could bring a more accurate and detailed calculus of Abdicating or exercising. In CHI, 2008.
the hyper-rectangles, a study to determine the best number [12] N. Liu, S. Lai, and C. Chen. Adaptive music
of songs is also planned. recommendation based on user behavior in time slot.
Journal of Computer Science and Network Security,
In summary, we can report that Improvise suggests songs 9(2):219–227, 2009.
suitable for daily activities to users with di↵erent tastes [13] D. McEnnis and S. J. Cunningham. Sociology and
and preferences. Moreover, the proposed model is flexible music recommendation systems. In Proceedings of the
enough to constantly adapt its recommendations accord- 8th International Conference on Music Information
ingly to the user feedback, while providing an answer to Retrieval, pages 185–186, 2007.
the cold-start problem.
[14] S. E. Park, S. Lee, and S.-g. Lee. Session-Based
Collaborative Filtering for Predicting the Next Song.
6. ACKNOWLEDGMENTS First ACIS/JNU International Conference on
This work was supported by national funds through Fun- Computers, Networks, Systems and Industrial
dação para a Ciência e Tecnologia, under INESC-ID multi- Engineering, 2011.
annual funding - PEst-OE/EEI/LA0021/2013 and LaSIGE [15] S. Reddy and J. Mascia. Lifetrak: Music in tune with
Strategic Project - PEst-OE/EEI/UI0408/2014. Ricardo your life. In Proceedings of the 1st ACM International
Dias was supported by FCT, grant reference SFRH/BD/ Workshop on Human-centered Multimedia, 2006.
70939/2010. [16] M. Schedl and P. Knees. Personalization in
Multimodal Music Retrieval. In 9th International
7. REFERENCES Workshop on Adaptive Multimedia Retrieval, 2011.
[1] G. D. Abowd, A. K. Dey, P. J. Brown, N. Davies, [17] B. Schwartz. The Paradox of Choice: Why More Is
M. Smith, and P. Steggles. Towards a better Less. Harper Perennial, 2005.
understanding of context and context-awareness. In [18] B. Shao, M. Ogihara, D. Wang, and T. Li. Music
Proc. of the 1st international symposium on Handheld recommendation based on acoustic features and user
and Ubiquitous Computing, 1999. access patterns. Audio, Speech, and Language
[2] L. Baltrunas, X. Amatriain, and V. Augusta. Towards Processing, IEEE Transactions on, 17(8):1602 –1611,
Time-Dependant Recommendation based on Implicit nov. 2009.
Feedback. In Context Aware Recommendation Systems [19] J. Su, H. Yeh, and P. Yu. Music recommendation using
(CARS - RecSys), 2009. content and context information mining. Intelligent
[3] R. Cai, C. Zhang, C. Wang, L. Zhang, and W.-Y. Ma. Systems, IEEE, (IEEE Intelligent Systems), 2010.
Musicsense: contextual music recommendation using [20] K. Verbert, D. Parra, P. Brusilovsky, and E. Duval.
emotional allocation modeling. In Proceedings of the Visualizing recommendations to support exploration,
15th international conference on Multimedia, transparency and controllability. In Proceedings of the
MULTIMEDIA ’07, pages 553–556, New York, NY, 2013 International Conference on Intelligent User
USA, 2007. ACM. Interfaces, 2013.
[4] P. Cano, M. Koppenberger, and N. Wack. An [21] X. Wang, D. Rosenblum, and Y. Wang. Context-aware
industrial-strength content-based music mobile music recommendation for daily activities. In
recommendation system. In Proceedings of the 28th Proc. of the 20th international conference on
annual international ACM SIGIR conference on Multimedia, 2012.
Research and development in Information Retrieval, [22] B. Yapriady and A. L. Uitdenbogerd. Combining
2005. demographic data with collaborative filtering for
[5] O. Celma. Music Recommendation and Discovery. automatic music recommendation. Springer, 2005.
Springer, 1st edition, 2010. [23] E. Zheleva, J. Guiver, E. Mendes Rodrigues, and
[6] W. W. Cohen and W. Fan. Web-collaborative N. Milić-Frayling. Statistical models of music-listening
filtering: recommending music by crawling the Web. sessions in social media. Proceedings of the 19th
Computer Networks, 33, 1999. international conference on World wide web - WWW
[7] R. Dias and M. J. Fonseca. Improving music ’10, 2010.
recommendation in session-based collaborative
filtering by using temporal context. In ICTAI, 2013.
[8] M. Hall, E. Frank, G. Holmes, B. Pfahringer,
P. Reutemann, and I. H. Witten. The weka data
mining software: An update. SIGKDD Explor. Newsl.,
2009.
[9] B. P. Knijnenburg, S. Bostandjiev, J. O’Donovan, and
A. Kobsa. Inspectability and control in social
recommenders. In Proceedings of the Sixth ACM
32