DJ-Running: An Emotion-based System for Recommending Spotify

Songs to Runners

Alvarez

1 a

, A. Guiu

, J. R. Beltr

2 b

, J. Garc

ıa de Quir

and S. Baldassarri

1 c

Department of Computer Science and Systems Engineering, University of Zaragoza, Zaragoza, Spain

Department Electronic Engineering and Communications, University of Zaragoza, Zaragoza, Spain

Keywords:

Running, Music Recommendations, Runners’ Emotions, Motivation and Performance.

Abstract:

People that practice running use to listen to music during their training sessions. Music can have a positive

inﬂuence on runners’ motivation and performance, but it requires selecting the most suitable song at each

moment. Most of the music recommendation systems combine users’ preferences and context-aware factors

to predict the next song. In this paper, we include runners’ emotions as part of these decisions. This fact has

forced us to emotionally annotate the songs available in the system, to monitor runners’ emotional state and to

interpret these data in the recommendation algorithms. A new next-song recommendation system and a mobile

application able to play the recommended music from the Spotify streaming service have been developed. The

solution combines artiﬁcial intelligence techniques with Web service ecosystems, providing an innovative

emotion-based approach.

1 INTRODUCTION

Most of the people that practice running as a sport,

listen to music during the development of the physi-

cal activity using their portable music players, mainly,

mobile phones. Running with music can help to in-

crease the runner’s motivation, making hard training

sessions much more pleasant as well as make the run-

ner feel less alone. These effects are of special inter-

est for long-distance runners, or even people with a

sedentary lifestyle who wish to start running. Since

music has the ability to produce reactions in the in-

dividuals (Terry, 2006), it cannot be ruled out that

the listening of different types of music could also

inﬂuence the motivation and performance of a run-

ner (Brooks and Brooks, 2010). Therefore, a runner

should select ideally a different playlist for each ses-

sion depending on the type of training, his/her musi-

cal preferences, his/her mood, or the characteristics of

the route he/she wants to run.

Music Recommendation Systems (MRS) reduce

the effort of users for selecting songs by consider-

ing their proﬁles and preferences. Traditionally, these

systems have been used to create personalised playlist

https://orcid.org/0000-0002-6584-7259

https://orcid.org/0000-0002-7500-4650

https://orcid.org/0000-0002-9315-6391

that were then played on portables music players.

The rise of music streaming services has motivated

a new generation of recommendation systems more

dynamic and ﬂexible than traditional approaches, for

example, next-song recommendation systems (Zheng

et al., 2019; Baker et al., 2019; Vall et al., 2019b).

These MRS create playlists in real time, while the

user is listening them, and learn from users’ short-

term and long-term behaviors for improving their pre-

dictions. Additionally, the widespread use of mobile

phones as players allows these systems to collect data

about the user’s context and utilize this contextual

information to better satisfy users’ interests (Kamin-

skas and Ricci, 2011; Wang et al., 2012; Chen et al.,

2019). Despite the advances in developing context-

aware systems, certain contextual decision factors are

still underused, such as users’ emotions and the ac-

tivity that they are carrying out at each moment, for

instance.

DJ-Running is a research project that monitors the

runners’ emotional and physiological activity during

the training sessions, to automatically recognize their

feelings and to select, in real-time, the most suitable

music to improve their motivation and performance

(DJR, 2019). In this paper, we present the context-

aware music recommendation system developed as

part of the project. It predicts the next song to be

played considering the user’s location and emotions

Álvarez, P., Guiu, A., Beltrán, J., García de Quirós, J. and Baldassarri, S.

DJ-Running: An Emotion-based System for Recommending Spotify Songs to Runners.

DOI: 10.5220/0008164100550063

In Proceedings of the 7th International Conference on Sport Sciences Research and Technology Support (icSPORTS 2019), pages 55-63

ISBN: 978-989-758-383-4

 2019 by SCITEPRESS – Science and Technology Publications, Lda. All rights reser ved

and the type of training session that he/she is carry-

ing out. This combination of contextual factors is a

contribution with respect to existing next-song rec-

ommendation proposals. In addition to the system,

we also present the mobile application that monitors

the runner’s activity and connects with the Spotify mu-

sic streaming service to play the recommended songs.

From a technological point of view, the solution has

been programmed combining the paradigms of cloud,

fog and service computing.

The remainder of this paper is organized as fol-

lows. In Section 2, the related literature regarding to

next-song recommendation systems is reviewed, pay-

ing attention in those solutions that are integrated with

the Spotify services. Sections 3 and 4 describe the

functionality and architecture of the DJ-Running ap-

plication and recommendation system, respectively.

The system consists of a set of Web services that

are detailed in Sections 5 and 6. Finally, Section 7

presents the paper’s conclusions and makes sugges-

tions for future research.

2 RELATED WORKS

2.1 Next-song Recommendation

Systems

The task of this class of recommendation systems is

to predict the next song that a user would like to listen

to. These predictions are mainly based on different

factors related to users and songs (Knees et al., 2019).

From the users’ point of view, factors can be classi-

ﬁed as intrinsic or external. The former are related

to users’ proﬁle and preferences, listening histories

and cultural features; whereas the external are mainly

associated with the user’s context, including the loca-

tion (Kaminskas and Ricci, 2011; Su et al., 2010), the

social context (Chen et al., 2019), the activity (Wang

et al., 2012; Oliver and Kreger-Stickles, 2006) or the

mood (Han et al., 2010), for instance. These con-

textual factors are more dynamics and have recently

gained relevance thanks to users listen to music on

their smart mobile phones. Today, these devices in-

tegrate rich sensing capabilities that allow to collect

a wide variety of data about users’ context and, as a

result, most of the current recommendation systems

make context-aware predictions (Wang et al., 2012).

On the other hand, songs’ features have been also

used by these systems (Cano et al., 2005; Borges and

Queiroz, 2017; Vall et al., 2019a; Zheng et al., 2019),

specially songs’ audio features. The interest in these

approaches lies in their recommendations are based

on data that can be extracted from songs and, there-

fore, are directly available. This availability alleviates

the cold start problem, which arises when there is not

data about a new user and, therefore, it is not possi-

ble to make effective recommendations (Chou et al.,

2016).

Runners’ emotions and types of training sessions

are two factors considered as part of the DJ-Running

recommendation system. In the ﬁeld of affective com-

puting many works propose methods to determine the

emotions perceived by the user while listens to a song

(Yang et al., 2018). These methods enhance songs’

metadata by adding them emotional labels. Never-

theless, next-song recommendations systems do not

use these labels to improve their results according to

users’ mood, for instance. Alternatively, the activity

that the user is performing at that instant is another

relevant factor, as it was discussed in (North et al.,

2004; Levitin et al., 2007). Nevertheless, it is usually

ignored by recommendation systems. As an excep-

tion, (Wang et al., 2012) proposes a solution that in-

fers automatically a user’s activity (working, sleeping,

studying, running, etc.) and recommends songs suit-

able for that activity, and (Oliver and Kreger-Stickles,

2006) creates playlists to help users to achieve their

exercise goals (walking or jogging, for example).

The techniques used to recommend the next song

are varied, even most of them are those commonly

used in general purpose recommendation systems:

the collaborative ﬁltering (Lee et al., 2010; Vall

et al., 2019a; Baker et al., 2019; Chen et al., 2019),

the content-based ﬁltering methods (Pampalk et al.,

2005; Cano et al., 2005; Oliver and Kreger-Stickles,

2006) and the hybrid approaches that combine both

techniques to mitigate their possible drawbacks (Vall

et al., 2019b). The collaborative techniques focus on

the analysis of the similarity and the relationships be-

tween users. It is becoming more frequent that the

friendship between social network users is included

in the models of these collaborative approaches (Chen

et al., 2019). Alternatively, the content-based meth-

ods are mainly based on songs’ features and users’

preferences and contexts. Nevertheless, nowadays the

tendency is to integrate different techniques to take

advantage of the strengths of each of them. These hy-

brid proposals also combine the common recommen-

dation techniques with probabilistic models (Wang

et al., 2012; Borges and Queiroz, 2017) or different

classes of neuronal networks (Choudhary and Agar-

wal, 2017; Zheng et al., 2019), for instance.

On the other hand, there is a growing interest in

solving the problem of automating the music playlist

continuation (Oliver and Kreger-Stickles, 2006; Vall

et al., 2019a; Baker et al., 2019). It consists of pro-

icSPORTS 2019 - 7th International Conference on Sport Sciences Research and Technology Support

viding a personalised extension to the playlist that the

user is listening to. Therefore, it is a particular case

of next-song approaches in which the recommenda-

tions are made by considering the songs previously

played (each playlist is processed as a user’s listen-

ing history). The factors that are usually involved in

these recommendations are songs’ order and popu-

larity (Vall et al., 2019b), the songs’ audio features

(Vall et al., 2019a; Baker et al., 2019), the most lis-

tened authors and musical genres, or users’ response

to recommended songs (Pampalk et al., 2005; Oliver

and Kreger-Stickles, 2006; Choudhary and Agarwal,

2017), among others. This last factor gains rele-

vance in these recommendation systems as an imme-

diate and explicit feedback that helps to maximize the

user’s satisfaction. This feedback can be determined

by using sensors to detect physical and physiologi-

cal responses (Oliver and Kreger-Stickles, 2006), by

evaluating the effects of playing the same list of songs

in a different order (Choudhary and Agarwal, 2017),

or by analyzing the user’s dislikes (for example, the

songs that are skipped by the user (Pampalk et al.,

2005)).

2.2 Music Recommendation Systems

based on Spotify

Most of the works related to Spotify propose mu-

sic recommendation systems to help users to create

their playlists. Recommendations are based on the

user’s preferences (musical genres and artists, mainly)

and the features of songs that he/she usually lis-

tens to. Users’ proﬁles are determined by utilizing

users’ past interactions (Fessahaye et al., 2019; Ben-

net, 2018) or by processing the messages published

by users in social networks, such as Twitter (Pichl

et al., 2015) or Facebook (Germain and Chakareski,

2013). Internally, these recommendation systems are

programmed integrating content and collaborative ﬁl-

tering techniques (Fessahaye et al., 2019; Pichl et al.,

2015; Germain and Chakareski, 2013). The ﬁrst ones

help to determine the similarity between songs based

on their audio features, while the latter determine the

similarity between users based on their preferences.

The same approach is currently used by Spotify (Ma-

dathil, 2017). Exceptionally, (Bennet, 2018) makes

recommendations using clustering techniques. As a

conclusion, with the exception of (Fessahaye et al.,

2019), these Spotify-based systems do not consider

music emotions. (Fessahaye et al., 2019) adds a mood

value to the set of songs’ features which is obtained

from the Million Playlist Dataset, released by Spotify

in 2018 as part of the Spotify RecSys Challenge.

3 THE DJ-Running APPLICATION

The DJ-Running technological infrastructure allows

a runner to conﬁgure his/her proﬁle in order to listen

personalized music during the training sessions. The

runner accesses this functionality by a mobile appli-

cation that works with the Spotify streaming service

to play the recommended songs. The aim of this sec-

tion is to describe the design of this application and

its interaction with the user and the environment.

Before using the mobile application for the ﬁrst

time, the runner sets up his/her personal proﬁle

through a Web application. This proﬁle includes an-

thropometric characteristics, demographic data, mu-

sical preferences, the type of runner, etc. It is highly

recommended to update this information periodically.

On the other hand, the user also has to purchase a li-

cense of Spotify. When the application is installed in

the mobile phone the ﬁrst time, it requires the user’s

credentials created by the Web application (to access

to the information of the proﬁle) and the license code

of Spotify. Once the software installation is com-

pleted, the application is ready to be used during the

training sessions. Figure 1-a shows the initial appli-

cation screen. Before starting to run, the runner must

introduce the kind of training session that he/she will

do and his/her emotional state at that moment (happy,

relax, stressed, etc.). The screen in which the runner

is asked about the emotional state is presented in Fig-

ure 1-b, and it is based in the ’Pick-A-Mood’ (PAM)

model (Desmet et al., 2012), a cartoon-based pictorial

instrument for reporting and expressing moods. This

model classiﬁes the possible emotional states of the

user according to the reference model of affect pro-

posed by Russell (Russell, 1980). Optionally, the ap-

plication can also be connected to an emotional wear-

able developed in the frame of our project (

Alvarez

et al., 2019). This device incorporates a set of sen-

sors (GSR, HR and oximeter) that allow to detect the

runner’s emotional state during the training session.

These estimations are carried out by an artiﬁcial in-

telligent system that translates the low level signals

of the sensors into emotions. After this initial con-

ﬁguration stage, the application starts to play person-

alized music by the streaming of Spotify. Figure 1-c

shows the interface of the application while it is play-

ing the music. The runner can stop the player, skip

the song (this actions is considered for future recom-

mendations) o stop the training session (top right of

the screen). During this stage, the application also

monitors the runner’s geographic location.

The decision about which song must be played at

each moment is complex. The DJ-Running recom-

mendation service is in charge of taking this deci-

DJ-Running: An Emotion-based System for Recommending Spotify Songs to Runners

Figure 1: Interface of the DJ-Running mobile application.

sion based in three different types of input parameters:

the runner’s proﬁle (set when he/she registered in the

system), the data explicitly introduced by the runner

before starting the training session (kind of training

and emotional state), and, lastly, the data automati-

cally recorded during the sport activity (runner loca-

tion, his/her heart rate, running pace and the changes

experienced in his/her emotional state, mainly). Fur-

thermore, the recommendation system also interacts

with the Spotify services to access to the musical pref-

erences of the user, and with different geographic

systems that offer relevant information related with

the runner’s current location (meteorological informa-

tion, level of noise, kind of terrain, altimetry, etc.).

The system takes into account all this information to

determine, in a personalized way, the next song to

play. The technological components that are involved

in this complex process are described in detail in the

following section.

Finally, when the runner ends the training session,

the application shows in a Google Maps map the route

followed and a summary of the session activity.

4 ARCHITECTURE AND

DEPLOYMENT OF THE

DJ-Running SYSTEM

Figure 2 shows the software architecture of the DJ-

Running system. Once the runner has signed up into

the system, he/she connects with the music recom-

mendation service using the DJ-Running mobile ap-

plication (right part of the ﬁgure). Internally, the ar-

chitecture consists of a set of Web services that pro-

vide the functionality needed to create users’ proﬁles,

to access music information and geographic data,

and to make personalised music recommendations.

These services require to interact with other external

providers to offer the mentioned functionality, for ex-

ample, with Spotify, Acoustic-Brainz, Weather Under-

ground or some Spatial Data Infrastructures (left part

of the ﬁgure).

The Music recommendation service is the core of

the DJ-Running system. It receives the recommenda-

tion request from the runner’s mobile application and

predicts the next song to play. These recommenda-

tions are made considering the runner’s context infor-

mation, the runner’s proﬁle and emotionally-labelled

songs available in the system. Most of these data are

accessible through the DJ-Running data services: the

User Proﬁle service, the Geo Data service and the

Music Delivery service. On the other hand, the Music

data service integrates a Music Emotion Recognition

systems (MER) that has been used to annotate emo-

tionally Spotify songs. A complete description of all

these services is presented in the next section.

From a technological point of view, the Spring

framework has been used for developing all these ser-

vices (Pivotal Software, 2017). It provides the core

features needed for programming, conﬁguring and de-

ploying any Java-based service. More speciﬁcally,

these services have been implemented as RESTful

Web services in order to facilitate their integration

in Web-accessible applications. On the other hand,

the Kubernetes technology has been also used to au-

icSPORTS 2019 - 7th International Conference on Sport Sciences Research and Technology Support

Figure 2: Architecture of the DJ-Running system.

tomate and manage the deployment of services in

cloud-based execution environments. Ideally, these

deployments should consider runners’ location for re-

ducing the communication latency and improving the

quality of service, in particular the deployment of the

music recommendation service. The fog computing

paradigm (Mahmud et al., 2018) has been applied to

fulﬁll these deploy requirements.

Figure 3: Deployment of the DJ-Running services.

Figure 3 shows the deployment of DJ-Running

system. Data Web services have been deployed and

executed in a public cloud provider, more speciﬁcally,

in the Amazon EC2 infrastructure. These services are

the back-end of the system and do not interact di-

rectly with runners and their applications. The mu-

sic recommendation service is the front-end of the

system and, therefore, it must reside at the edge of

the network (or fog node). We have programmed the

recommendation system as a lightweight service that

can be deployed into a node with low computing and

storage capacity, for example, into a Raspberry Pi 3

computer. In our deployment, this small-size com-

puter has been connected to a network close to the

area where the runners who participate in the system

test-beds ran. Additionally, different instances of the

recommendation service could be easily conﬁgured

and deployed into geographically distributed servers

to cover a wider geographical area.

5 WEB DATA SERVICES

In this section the functionality of the Web data ser-

vices and the most relevant implementation decisions

are detailed.

5.1 The User Proﬁle Service

An user model has been deﬁned to create a digital

representation of runners. It includes runners’ anthro-

pometric characteristics (height, weight, etc.), demo-

graphics data (age, genre, education, etc.), musical

preferences (musical genres, authors, popularity, etc.)

and social relationships. Most of these data are cur-

rently completed by the runner as part of the sign-up

process, and can be updated via the Web application at

any moment. Others can be automatically gathered by

the system. For example, musical preferences can be

continuously updated from the user’s Spotify listen-

ing histories, or friendships can be determined from

the user’s social network account (such as, Twitter or

Facebook). Both gathering processes require the user’

extra information and permissions. On the other hand,

the model also includes the songs that are skipped by

an runner as part of the musical preferences. This

feedback is sent to the system by the runner’s mobile

application during training sessions.

The User Proﬁle service stores runners’ descrip-

tions deﬁned from the previous model and manages

DJ-Running: An Emotion-based System for Recommending Spotify Songs to Runners

the gathering processes needed to update them. Its

Web API provides access to these descriptions, offer-

ing a set of search operations. This functionality is

mainly used by the music recommendation service.

5.2 Services for Accessing to Music

Information

Two services are responsible for managing and pro-

viding access to the collection of songs available in

the DJ-Running system. The Music Data service in-

teracts with the Spotify Web API for developers in

order to retrieval songs of interests by applying dif-

ferent criteria (songs’ popularity, musical genres, or

artists, for instance). Metadata and audio features of

these songs are requested to the music provider and

stored them locally into the service’s database. Sub-

sequently, these songs are annotated emotionally by

a Music Emotion Recognition system (MER). It inte-

grates a set of machine learning models able to deter-

mine the emotion perceived by an user when listening

to a song from its audio features. These emotions are

represented by means of labels that are stored jointly

with song’s metadata into the service‘s database. The

current version of our annotated database contains

over 60,000 popular songs. Nevertheless, the sys-

tem has been developed for automatically processing

a massive collection of songs by providing an alter-

native to other approaches based on evaluations with

users or experts.

In this work, the emotional labels are based

on the Russell’s circumplex model (Russell, 1980),

one of the most popular dimensional models in af-

fective computing. It represents affective states

over a two-dimensional space that is deﬁned by

the valence (X-axis) and arousal (Y-axis) dimen-

sions. The valence represents the intrinsic plea-

sure/displeasure (positive/negative) of an event, ob-

ject or situation, and the arousal the feeling’s in-

tensity. The combination of these two dimensions

(valence/arousal) determines four different quadrants:

the aggressive-angry (negative/positive), the happy

(positive/positive), the sad (negative/negative) and the

relaxed (positive/negative) quadrant. In our proposal

the emotional annotation of a song can have one of

these four values: Angry, Happy, Sad and Relaxed.

This annotation represents the emotion perceived by

the users and corresponds with one of the Russell

model’s quadrants.

On the other hand, the Music Delivery service pro-

vides the recommendation system with access to the

songs stored in the database. Internally, it creates a

set of indexes to improve the efﬁciency of searches in

large-scale datasets of emotionally annotated songs.

These indexes were built considering the decision

rules that guide the recommendation process of songs.

Additionally, the service provides a wide variety of

search criteria and options to facilitate the access to

songs of interest.

5.3 The Geographic Data Service

Some geographic data related to the runner’s location

are used as contextual factors by the DJ-Running rec-

ommendation service. We are mainly interested in

meteorological information and data that describe the

environment in which the user is running. The Geo

Data Service is responsible for interacting with Web

services that provide these data as well as updating

periodically a local database in which the geodata of

interest are stored. Some of these data are change-

able and, therefore, they must be constantly updated

(the meteorological, noise-level or pollution data, for

instance); whereas, others have a more static nature

(the type of terrain, the average slope of terrain, the

altitude, etc.).

Currently, the service interacts with the Weather

Underground API to get meteorological information

at real time. Besides, it integrates the set of services

published by the Spanish Spatial Data Infrastructure

(SDI). These enable the download and analysis of a

wide variety of geographic data produced in Spain.

6 MUSIC RECOMMENDATION

SERVICE

Figure 4 shows the high-level architecture of the rec-

ommendation service. The core component is the

recommendation system. Its implementation is based

on a Nearest Neighbor Search algorithm (NNS). This

class of algorithms solve the problem of ﬁnding the

point in a given set that is closest (or most similar) to

a given point. Formally, they are deﬁned from a set of

points in a space M and a metric distance that allows

to determine the similarity (or dissimilarity) between

these points.

In our proposal, each Spotify song has been trans-

lated to a point of the space M. A point is a nu-

meric vector created from the song’s audio features

and emotional labels. Before starting the service exe-

cution, these points have been calculated from songs

available into the Music Delivery service, and stored

them into a local repository (it represents the space

M). Different indexes have been also created to im-

prove the performance of executing the metric dis-

tance on the repository of points.

icSPORTS 2019 - 7th International Conference on Sport Sciences Research and Technology Support

Figure 4: DJ-Running Music Recommendation system.

Despite these indexes, the search space of our

problem is complex (Spotify stores more than 30 mil-

lion of songs and a point is deﬁned for each of these

songs). Besides, it is not necessary to retrieve an exact

search result, which is an overkill in this type of appli-

cations. Therefore, we have decided to use the Annoy

algorithm (Erik Bernhardsson, 2013), an approximate

NNS algorithm that has provided good results in the

ﬁeld of multimedia recommendations. These algo-

rithms are able to retrieve approximate nearest neigh-

bors much faster than NNS algorithms. Internally,

our implementation of Annoy was conﬁgured for us-

ing the angular distance as similarity measurement

between the points. Its performance and accuracy

were evaluated by executing the ANN-benchmarks en-

vironment, a benchmarking framework for approxi-

mate nearest neighbor algorithms search (Aum

uller

et al., 2018), by obtaining satisfactory results from the

point of view of our problem.

The other two components of the recommenda-

tion system are the Input analyzer and a set of Fil-

ters. The former is responsible for determining what

kind of song (or songs) will be recommended to the

runner considering the service request’s input param-

eters (the runner location and mood, the kind of train-

ing, the emotion to be induced to the runner, or level

of fatigue, among others). A rule engine translates

these parameters to a vector that describes the audio

and emotional features of the song (or songs) to ﬁnd

in the search space. The emotional labels of interest

are mainly determined from the geographic and envi-

ronmental data related to the runner location and the

emotions that the system tries to induce him/her. Fi-

nally, the resulting vector corresponds with the point

in the space that will be submitted as input to the rec-

ommendation system.

On the other hand, a list of candidate songs is re-

turned as output of a recommendation request. Cur-

rently, 50 songs are recommended for each input re-

quest (this number is easily conﬁgured). Filters are

responsible for scoring and ranking the returned songs

in order to personalize the ﬁnal recommendations.

Two class of ﬁlters have been implemented: Per-

sonal ﬁlters and Social ﬁlters. The ﬁrst ones score

the candidate songs in accordance with the runner’s

musical preferences. These preferences were speci-

ﬁed by the runner when he/she registered in the sys-

tem. The social ﬁlters are based on the concept of

similarity between runners. This similarity has differ-

ent dimensions: the range of age groups, the physi-

cal/emotional response to a kind of training, the mu-

sical preferences or the friendship in social networks

(such as Twitter or Facebook), for instance. Cluster-

ing algorithms and Collaborative ﬁltering techniques

have been combined to implement a prototype version

of these social ﬁlters. Once all these ﬁlters have been

executed, the three best ranked songs are returned as

result of the recommendation service.

Finally, as an example, we explain the kind of

songs that would be recommended for a speciﬁc train-

ing session. The runner’s proﬁle is a man, middle-

aged and half-marathon runner. He is in the ﬁnal stage

training previous to a race and, therefore, he needs to

maintain a good running pace. The training session

consists of three parts: 15 minutes easy run, then 10 x

(1 minute fast, 1 minute jog) and, ﬁnally, 15 minutes

easy run (a total of 50 minutes). Besides, the route

is mostly ﬂat to favour a good running pace, and the

day is sunny and with a pleasant temperature. In this

example, if the runner is having a good performance

during the training, the input analyzer will propose the

following kinds of songs. For the ﬁrst part of the train-

ing, there are selected happy songs, but not specially

motivating. Then, motivating and relaxing songs will

be alternated for each of the 10 repetitions (for the fast

and jog pace, respectively). And, ﬁnally, calm and re-

laxed songs will be again selected to make easier the

runner’s recovery during the last part of the session.

Nevertheless, if the runner’s performance is not as ex-

pected, alternative songs should be proposed (for ex-

ample, motivating songs that help him to improve an

undesirable running pace during the third part of the

training as an alternative to calm and relaxed ones).

In any of the cases, these emotional labels are deter-

DJ-Running: An Emotion-based System for Recommending Spotify Songs to Runners

mined by the analyzer at real-time and, later, used by

the recommendation algorithm to determine the con-

crete songs to be played at each moment. As conclu-

sion, these songs have been selected by considering

the runner’s proﬁle, location and emotional state.

7 CONCLUSIONS AND FUTURE

WORK

In this paper a next-song recommendation system for

runners has been proposed. The system makes per-

sonalised recommendations to increase runners’ mo-

tivation and performance. Different context-aware

factors are considered as part of the recommendation

process, specially, the runner’s emotions and the kind

of training session. The inclusion of users’ emotions

in this process has forced us to emotionally annotate

Spotify songs using machine-learning techniques and,

then, to interpret these annotations in the recommen-

dation algorithms. The current version of the system

consists of a mobile application able to play music

from the Spotify streaming service and a set of Web

services that provide the functionality needed to make

context-aware recommendations.

The DJ-Running system is being currently vali-

dated with real users (more speciﬁcally, with triath-

letes) in collaboration with the Sports Medicine Cen-

tre of the Government of Arag

on (Spain). We are in-

terested in studying the runners’ emotional and phys-

ical responses when they train with certain kind of

music. Triathletes must complete the session of 50

minutes described in the previous section on two oc-

casions (during two consecutive weeks, once a week).

The ﬁrst time, the system is conﬁgured for recom-

mending songs that produce a positive effect in the

runner. We have used the emotional labels described

in the example for selecting the songs to be played.

The second time, the goal is to produce a negative

effect in the runner and, therefore, we have selected

noisy and aggressive-angry songs instead of the ones

proposed previously. Although this validation is still

in progress, preliminary results show that the differ-

ent kinds of songs affect runners’ performance and

perceive exertion.

As future work, validation results will be used to

improve the interpretation of context-aware factors as

well as the recommendations. Besides, we would

like to enhance the proposed user model, to automate

the processes for gathering registered users’ data, and

to improve the efﬁciency of recommendation algo-

rithms. Our ﬁnal aim is to publish a version of our

application on Google Play.

ACKNOWLEDGEMENTS

This work has been supported by the TIN2015-

72241-EXP project, granted by the Spanish

Ministerio de Econom

ıa y Competitividad, and

the DisCo-T21-17R and Affective-Lab-T25-17D

projects, granted by by the Aragonese Govern-

ment and the European Union through the FEDER

2014-2020 “Construyendo Europa desde Arag

on”

action.

REFERENCES

(2019). The DJ-Running project. https://djrunning.es/.

Alvarez, P., Beltr

an, J. R., and Baldassarri, S. (2019). Dj-

running: Wearables and emotions for improving run-

ning performance. In Ahram, T., Karwowski, W., and

Taiar, R., editors, Human Systems Engineering and

Design, pages 847–853, Cham. Springer International

Publishing.

Aum

uller, M., Bernhardsson, E., and Faithfull, A. (2018).

Ann-benchmarks: A benchmarking tool for approxi-

mate nearest neighbor algorithms.

Baker, J. C., Averill, Q. S., Coache, S. C., and McA-

teer, S. (2019). Song recommendation for automatic

playlistcontinuation. Digital WPI, Retrieved from

https://digitalcommons.wpi.edu/mqp-all/6744.

Bennet, J. (2018). A scalable recommender system for auto-

matic playlist continuation. Data Science Dissertation.

University of Sk

ovde, School of Informatics, Sweden.

Borges, R. and Queiroz, M. (2017). A probabilistic model

for recommending music based on acoustic features

and social data. In 16th Brazilian Symposium on Com-

puter Music, pages 7–12.

Brooks, K. and Brooks, K. (2010). Enhancing sports perfor-

mance through the use of music. Journal of exercise

physiology online, 13(2):52–58.

Cano, P., Koppenberger, M., and Wack, N. (2005). Content-

based music audio recommendation. In Proceedings

of the 13th annual ACM international conference on

Multimedia, pages 211–212. ACM.

Chen, J., Ying, P., and Zou, M. (2019). Improving mu-

sic recommendation by incorporating social inﬂuence.

Multimedia Tools and Applications, 78(3):2667–

2687.

Chou, S.-Y., Yang, Y.-H., Jang, J.-S. R., and Lin, Y.-C.

(2016). Addressing cold start for next-song recom-

mendation. In Proceedings of the 10th ACM Confer-

ence on Recommender Systems, pages 115–118.

Choudhary, A. and Agarwal, M. (2017). Music recommen-

dation using recurrent neural networks.

Desmet, P., Vastenburg, M., Bel, D., V., and Romero, N.

(2012). Pick-a-mood; development and application of

a pictorial mood-reporting instrument.

Erik Bernhardsson (2013). Annoy. https://github.com/

spotify/annoy.

icSPORTS 2019 - 7th International Conference on Sport Sciences Research and Technology Support

Fessahaye, F., P

erez, L., Zhan, T., Zhang, R., Fossier, C.,

Markarian, R., Chiu, C., Zhan, J., Gewali, L., and Oh,

P. (2019). T-recsys: A novel music recommendation

system using deep learning. In 2019 IEEE Interna-

tional Conference on Consumer Electronics (ICCE),

pages 1–6.

Germain, A. and Chakareski, J. (2013). Spotify me:

Facebook-assisted automatic playlist generation. In

2013 IEEE 15th International Workshop on Multime-

dia Signal Processing (MMSP), pages 25–28.

Han, B.-j., Rho, S., Jun, S., and Hwang, E. (2010). Music

emotion classiﬁcation and context-based music rec-

ommendation. Multimedia Tools and Applications,

47(3):433–460.

Kaminskas, M. and Ricci, F. (2011). Location-adapted

music recommendation using tags. In Proceedings

of the 19th International Conference on User Model-

ing, Adaption, and Personalization, UMAP’11, pages

183–194, Berlin, Heidelberg. Springer-Verlag.

Knees, P., Schedl, M., Ferwerda, B., and Laplante, A.

(2019). User awareness in music recommender sys-

tems. In Personalized Human-Computer Interaction.

Lee, S. K., Cho, Y. H., and Kim, S. H. (2010). Collabora-

tive ﬁltering with ordinal scale-based implicit ratings

for mobile music recommendations. Information Sci-

ences, 180(11):2142–2155.

Levitin, D., D James Mcgill, P., Daniel, D., and Levitin, J.

(2007). Life soundtracks: The uses of music in every-

day life’.

Madathil, M. (2017). Music recommendation system spo-

tify - collaborative ﬁltering. Reports in Computer Mu-

sic. Aachen University, Germany.

Mahmud, R., Kotagiri, R., and Buyya, R. (2018). Fog Com-

puting: A Taxonomy, Survey and Future Directions,

pages 103–130. Springer.

North, A. C., Hargreaves, D. J., and Hargreaves, J. J. (2004).

Uses of music in everyday life. Music Perception: An

Interdisciplinary Journal, 22(1):41–77.

Oliver, N. and Kreger-Stickles, L. (2006). Papa: Physiology

and purpose-aware automatic playlist generation. In

ISMIR, volume 2006, page 7th.

Pampalk, E., Pohle, T., and Widmer, G. (2005). Dynamic

playlist generation based on skipping behavior. In IS-

MIR, volume 5, pages 634–637.

Pichl, M., Zangerle, E., and Specht, G. (2015). Combin-

ing spotify and twitter data for generating a recent and

public dataset for music recommendation. In Proceed-

ings of the 26th GI-Workshop Grundlagen von Daten-

banken (GvDB 2014), Ritten, Italy.

Pivotal Software (2017). Spring framework. https://spring.

io/projects/spring-framework.

Russell, J. (1980). A circumplex model of affect. Journal of

personality and social psychology, 39(6):1161–1178.

Su, J., Yeh, H., Yu, P. S., and Tseng, V. S. (2010). Music

recommendation using content and context informa-

tion mining. IEEE Intelligent Systems, 25(1):16–26.

Terry, P. (2006). Psychophysical effects of music in sport

and exercise: an update on theory, research and appli-

cation. pages 415–419.

Vall, A., Dorfer, M., Eghbal-zadeh, H., Schedl, M., Burjor-

jee, K., and Widmer, G. (2019a). Feature-combination

hybrid recommender systems for automated music

playlist continuation. User Modeling and User-

Adapted Interaction, 29:527–572.

Vall, A., Quadrana, M., Schedl, M., and Widmer, G.

(2019b). Order, context and popularity bias in next-

song recommendations. International Journal of Mul-

timedia Information Retrieval, 8(2):101–113.

Wang, X., Rosenblum, D., and Wang, Y. (2012). Context-

aware mobile music recommendation for daily activ-

ities. In Proceedings of the 20th ACM international

conference on Multimedia, pages 99–108. ACM.

Yang, X., Dong, Y., and Li, J. (2018). Review of data

features-based music emotion recognition methods.

Multimedia Syst., 24(4):365–389.

Zheng, H.-T., Chen, J.-Y., Liang, N., Sangaiah, A. K., Jiang,

Y., and Zhao, C.-Z. (2019). A deep temporal neu-

ral music recommendation model utilizing music and

user metadata. Applied Sciences, 9(4).

DJ-Running: An Emotion-based System for Recommending Spotify Songs to Runners