Social Tracks: Recommender System for Multiple Individuals using

Social Inﬂuence

Lesly Alejandra Gonzalez Camacho

, Jo

ao Henrique Kersul Faria,

Solange Nice Alves-Souza

and Lucia Vilela Leite Filgueiras

Departamento de Engenharia de Computac¸

ao e Sistemas Digitais,

Escola Polit

ecnica da Universidade de S

ao Paulo, S

ao Paulo, Brasil

Keywords:

Recommender System, Group Recommender System, Social Network, Social Inﬂuence.

Abstract:

The number of data generated through interactions within a social network, or interactions within a platform

resources (eg. clicks, hits, purchases), grow exponentially over time. The popularization of social networks

and the increase of interactions allow data to be analyzed to predict the tastes and desires of consumers. The

use of recommendation systems to ﬁlter content based on the characteristics and tastes of a user is already

widespread and applied across platforms. However, the application of recommendation systems to multiple

individuals is a less explored ﬁeld. For this project, data was gathered from social networks to recommend

music playlists to a group of individuals. Listening to music as a group is a common activity, be it with

friends, couples or in parties. Social network data are used to identify the social inﬂuence of the individuals

in the group. In addition, to identify the preferences, the characteristics of the songs most frequently heard

by the members of the group are assembled. Matrix factorization is used to predict group interests. Proposed

inﬂuence factor, based on similarity, leadership and expertise, is added to compute a ﬁnal recommendation. A

social network was created to support the controlled experiment, the results show the prediction made by the

system vary of 1,455 of the ratings made by the group’ members.

1 INTRODUCTION

Current applications increasingly use recommender

systems (RS) that suggest products and/or services of

interest to their users. Platforms, e-commerce and

social networks, such as Netﬂix, Spotify, Amazon,

and Facebook use RS to recommend content more

suited to the interests of individuals. The proposal of

an RS is to generate personalized recommendations

for each user using algorithms that evaluate the items

of their interest, based on identiﬁed preferences, as

well as data of other users with similar interests.

Recommendations based on a user’s

characteristics and preferences are already

widely disseminated and applied across platforms

(Contratres et al., 2018). However, activities such

as traveling, playing, watching movies, or listening

to music, can and are often performed by groups of

people. In these situations, an RS should go beyond

individual ratings and evaluate the preferences of

https://orcid.org/0000-0002-9387-8351

https://orcid.org/0000-0002-6112-3536

https://orcid.org/0000-0003-3791-6269

the group so that the result of the recommendation

is satisfactory to all. This approach is known as the

Group Recommendation System (GRS) (Ricci et al.,

2011) and is one of the aspects little explored in RS.

Thus, GRS identiﬁes interests common to individuals

in the group to generate recommendations.

As examples of applications for GRS, or with

multiple individuals, one has to obtain a better

place for traveling in a group, TV that can adapt

its programming according to the people who are

watching it, songs to be played in a car with several

passengers, among others.

Connected phones and the popularization of social

networks increase the generation and daily data ﬂow

(Quirino et al., 2015). Social networks data is rich

in information about how individuals relate, form

groups, share common tastes, and inﬂuence each

other. Several papers have been published in recent

years using social networking data to improve product

and service recommendations (Zhang et al., 2013;

Zhao et al., 2016; Lian et al., 2016; Prando et al.,

2017; Contratres et al., 2018; Gonzalez-Camacho and

Alves-Souza, 2018).

Camacho, L., Faria, J., Alves-Souza, S. and Filgueiras, L.

Social Tracks: Recommender System for Multiple Individuals using Social Inﬂuence.

DOI: 10.5220/0008166503630371

In Proceedings of the 11th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2019), pages 363-371

ISBN: 978-989-758-382-7

363

We here propose a RS that, based on the deﬁnition

of the members of the group, uses data from a social

network to identify social inﬂuence among these

individuals. An experiment, which includes a mobile

application that interfaces with the proposed RS, is

presented to validate the proposal. This application

is used to choose the group of friends and, in the

end, to present the result of the recommendation.

From the mobile application, the social network data

and the individual musical preferences of a musical

platform are recovered. Based on the calculated social

inﬂuence, a subset of the set of songs formed by

every-one’s preference is recommended to the group.

The social inﬂuence is calculated based on four

factors and has the greatest weight in identifying the

songs recommended to the group.

This work is divided as follows; Section

II provides details about the characteristics and

techniques employed in RS and GRS, besides

discussing the data of social networks and their

use in RS. Section III presents the RS proposed,

evidencing the calculation proposed to identify the

social inﬂuence on the group. In Section IV, the

RS architecture and details of the experiment are

presented. Section V; presents the results and Section

VI, the conclusion of the work.

2 RECOMMENDER SYSTEMS

In social relationships, it is natural for people to

recommend books, music, and movies to one another.

RS has come to assist and to extend this natural

process of content exchange (Resnick and Varian,

1997), is generally used when there are many items

to choose from and it is unfeasible for the user to be

aware of all the content available (Deng et al., 2014;

Lalwani et al., 2015; Al-Hassan et al., 2015)

There are different ﬁltering approaches employed

in RS, among which the most popular is Collaborative

Filtering (CF) (Ricci et al., 2011). CF aims to

calculate the similarity of a user in relation to others,

to recommend the active user items that other users

with similar likes preferred in the past.

CF is divided into two main models:

Neighborhood Methods and Latent Factor Models.

Neighborhood methods focus on calculating the

relationship between items, or between users, based

on the ratings made to the items by the active user, or

by their more similar neighbors (Koren et al., 2009).

In contrast to neighborhood methods, which use the

classiﬁcations stored directly in the forecast, Latent

Factor models use these classiﬁcations to learn a

predictive model. The idea is to be able to model

user-item interactions with factors that represent

latent characteristics of users and items, such as the

user preference class and the item category class

(Desrosiers and Karypis, 2011).

Latent factor models operate by characterizing

items and users in computationally inferred factors

based on user ratings. These factors are comparable

to the item categories, such as the genre of a movie

(Koren et al., 2009).

Matrix Factorization (MF) is a latent factor model

approach (Koren et al., 2009). In this approach,

each item i that can be recommended is described

by a vector q

of latent factors estimated from the

characteristics of the item. For each user u, there is

another vector p

that represents the user’s interest in

the items. Equation 1: estimates a user’s prediction

for an item.

= q

(1)

In computing terms, the highest processing cost

occurs in the calculation of q

and p

and an

efﬁcient method to do this is the Alternating Least

Squares (ALS) (Koren et al., 2009; Koren and Bell,

2011), which allows calculations to occur in parallel

since the parameters of the matrix are calculated

independently. In addition, ALS is the most efﬁcient

in systems based on implicit data (Hu et al., 2008).

In music recommendations, one does not always have

explicit data of the user’s rating for a particular song,

but it is implicitly inferred through the number of

times he has heard it.

2.1 Aggregation Techniques

Aggregation techniques are used to combine

individual recommendations, generating a

recommendation that can satisfy a users group.

Some of the main aggregation techniques are

(Masthoff, 2011):

• Average: Gets the average of the rating for a given

item.

• Multiplicative: multiplies the rating of each

individual, obtaining a value for the rating of the

group.

• Vote by approval: counts the number of times an

item has been evaluated above a certain threshold.

• Minor dissatisfaction: consider the lowest rating

as the group’s assessment.

• Greater satisfaction: considers the highest rating

for the group rating.

• Average without dissatisfaction: performs the

average of the ratings, disregarding items with

ratings lower than a threshold value.

KDIR 2019 - 11th International Conference on Knowledge Discovery and Information Retrieval

364

Each aggregation strategy generates a group

assessment for a particular item or set of items. For

example, using the ”Average” technique, if an item is

rated by user A with grade 8, by B with 9, and by C

with 10, the rating of the group consisting of A, B,

and C is 9 among the three ratings).

2.2 Social Inﬂuence Features in the

Recommendation to Groups

Differently from individual recommendation, in a

group recommendation, other social interactions need

to be taken into account. For example, the perception

of the whole group could be affected in case the

recommendation is in any way embarrassing for one

of the group members, (Masthoff and Gatt, 2006).

Masthoff and Gatt (2006) explains that the

satisfaction of a group can depend on two causes:

(i) emotional contagion, in which the satisfaction

or dissatisfaction of some users can lead to the

satisfaction, or not, of others; (ii) compliance,

in which the opinion of others may inﬂuence a

user’s opinion, whether by normative or information

inﬂuence. Normative is when the individual, by

wanting to be part of a group, expresses an opinion

equal to that of the group, even if inwardly he/she

does not agree. Conversely, informational is when the

opinion of the individual, in fact, changes, because

he/she believes that the group is correct.

Wang and Lu (2014) discuss the concept of

inﬂuence in which a user may have greater power of

contagion and compliance in a group. Discovering the

user with the greatest inﬂuence in a group becomes

relevant, because satisfying this user, increases the

chances of satisfying the whole group. Zhu and

Huberman (2014) show how people’s choices are

affected by the recommendations of others.

Masthoff (2011) combines the aggregation

techniques with the mechanisms of social psychology,

ending in three possible strategies that can only be

implemented if RS has a feedback mechanism in

which user satisfaction can be measured in real time.

These strategies are:

• Strongly support the grumpier: This strategy

recommends the item that the least satisﬁed

person likes most.

• Weakly supportive: This strategy selects items

that are reasonably appreciated by the least

satisﬁed member (items rated 8/10 or above).

• With weights: This strategy adds weights to the

users, depending on their satisfaction, and thus,

uses the weights at the moment of performing the

aggregation.

3 THE RECOMMENDER

SYSTEM PROPOSED

The RS is accessed by the using a mobile application

designed to simulate interaction with e-commerce.

This application (detailed below) accesses the user’s

social network data and the data about the musical

preferences of the members of the group on a musical

platform. The group of friends and credentials of the

user in the social network are informed in the login to

access the application. This information is sent to the

RS, which is composed of the following modules:

• Social inﬂuence calculator: it calculates the

inﬂuence of each individual in relation to the

group based on the proposed inﬂuence model.

• Individual recommendation calculator: it

calculates the recommendation for each member

of the group using data obtained from the musical

platform.

• Aggregation techniques: it aggregates the

individual recommendations considering the

inﬂuence. This module uses the aggregation

techniques detailed in 2.1 with the result of

applying the proposed inﬂuence model to create

the recommendation for the group.

• Music recommendation: A subset of songs

computed as a result of the recommendation.

In the proposed RS the individual

recommendations are identiﬁed using CF,

implementing MF, or more speciﬁcally, ALS.

The ALS is used to infer user preference for a set of

items, using the implicit ratings of other users.

Masthoff (2011) details two experiments

conducted to assess which aggregation technique

made the most sense for a group, between ”Minor

Dissatisfaction”, ”Average” and ”Average without

dissatisfaction”. The authors concluded that,

although it is simple, the technique of ”Media

without dissatisfaction” generates good results.

Thus, in this work, the technique ”Media without

dissatisfaction” is used to aggregate individual

recommendations.

3.1 Social Inﬂuence Modeling

Guo et al. (2016) details the inﬂuence of an individual

on other people, showing how this inﬂuence can be

a key factor for the recommendation. The authors

use ﬁve factors as the basis for calculating an user’s

inﬂuence on a group:

• Expertise Factor: expresses a user’s knowledge of

a given topic. Generally, the opinion of experts is

more accepted than that of others.

Social Tracks: Recommender System for Multiple Individuals using Social Inﬂuence

365

• Susceptibility Factor: considers how much a user

is susceptible to the opinions and emotions from

others.

• Personality Factor: deﬁned by the individual’s

behavior pattern. Guo et al. (2016) determines

personality factor performing Thomas Kilmann

Instrument (TKI), which identiﬁes the behavioral

trend of an individual to deal with the other

members of the group.

• Intimacy Factor: Measures how much a group is

connected. The more the members of a group

are close, the more likely they are to accept the

opinions of the others.

• Similarity Factor: evaluates the degree of

similarity among group members. Obtained

through activities and information in common.

In the approach proposed by Guo et al. (2016),

inﬂuence factors are difﬁcult to automate. For

example, to deﬁne the Personality Factor the TKI test

is performed, which consists in making a series of

interviews with each member of the group. Albeit

interesting, in an actual application, the time to

take the test will certainly discourage an individual

from continuing into the environment to receive the

recommendation. That is why the factors proposed

by Guo et al. (2016) are impracticable for a RS by the

way these factors are calculated.

Differently from Guo et al. (2016), 3 factors

to deﬁne the social inﬂuence are proposed here to

determine the preference of groups of users. These

factors are calculated through data coming from

social networks only. These factors are the Expert

Factor, the Leader Factor, and the Similarity Factor.

They are normalized on a scale from 0 to 1 and they

are detailed as follows:

• Expert Factor: evaluates a user’s knowledge of a

speciﬁc topic. For music, it measures how much

a user hears and knows artists.

i j

∑

i=1

∑

i6= j, j=1

n f

+n f

+na

+nm

(2)

Where E

i j

is the element of matrix E that

measures the users expertise, nu is the number of

users in the group, n f is the number of followers

in the user’s music application, na is the number

of artists the user follows, and nm is the number

of songs the user has in his/her library.

Fexp

∑

j=1

i j

(3)

Where Fexp

is the Expert Factor for user i, nu

is the number of users in the group.

• Leader Factor: measures how much a user

represents the ﬁgure of a leader in a group. It is

measured by the repercussion of his/her posts in

relation to the other members of the group.

i j

∑

i=1

∑

i6= j, j=1

+tl

+tm

(4)

Where L

i j

is the element of matrix L for

evaluating the repercussion of a users’ posts, nu

is the number of users in the group, tl

is the total

number of Likes user i received, tm

is the total

number of mentions user i received. Analogous to

and tm

Flead

∑

j=1

i j

(5)

Where Flead

is the Leader Factor for user i and

nu is the number of users in the group.

• Similarity Factor: Measures how similar a user is

to the others in a group. This factor is calculated

by observing the activities in common among

users within the social network.

i j

∑

i=1

∑

i6= j, j=1

j,i

i, j

+tl

j,i

i, j

+ts

(6)

Where S

i j

is the element of matrix S that lists the

posts of interest and shared by each user. S

i j

calculated between each pair of users, nu is the

number of users in the group, l

j,i

is the number

of posts user j liked of user i, s

j,i

is the number

of posts that user j shared of user i. Analogous to

i, j

and s

i, j

Fsim

∑

j=1

i j

(7)

Where Fsim

is the Similarity Factor for user i

and nu is the number of users in the group and

i j

• Inﬂuence Factor: The inﬂuence factor is the

average of the three previously deﬁned factors and

which is used for aggregating of the individual

recommendations.

Fin f

= (Fexp

+ Flead

+ Fsim

)/3 (8)

Eq. 9 includes the Inﬂuence Factor to calculate

the individual recommendations.

in f luence

= r

i, j

∗ (1 + Fin f

) (9)

Where r

in f luence

is the rating of a user i

considering his/her inﬂuence, r

i, j

is a user i’s

KDIR 2019 - 11th International Conference on Knowledge Discovery and Information Retrieval

366

rating for a given item j and Fin f

is the inﬂuence

factor of i. Thus, the inﬂuence factor changes the

rating calculated by the ALS for each user.

Finally, the rating of an item by the group ( RG

g, j

)

(eq. 10) is the aggregation of individual ratings

considering his/her social inﬂuence.

g, j

∑

i=1

in f luence

(10)

Where nu is the number of users in the group

and r

in f luence

is the rating of a user i considering

his/her inﬂuence.

4 ARCHITECTURE AND

IMPLEMENTATION

Figure 1 shows the proposed RS, called Social Tracks,

architecture divided into three main parts:

Figure 1: The RS Architecture proposed.

1. Management API: makes the connection between

the mobile application and the recommendation

API. Here data are organized, formatted and sent

to the Recommender API. Later the results are

also organized and suitable to be shown in the

Mobile Application to the user.

2. Recommender API: processes users’ ratings for a

set of items as well as social network information

to generate a recommendation based on social

inﬂuence.

3. User Interface: allows the user to interact with

RS. Through this, implemented by a mobile

application, the user provides both data from

his/her music activity (Spotify) and from his/her

social network, creates and manages the groups,

receives the recommendations for a selected

group and can listen to and rate the list of

recommended songs.

4.1 Recommendation Data-ﬂow

Figure 2 details the recommendation process and the

type of information generated in each step. The

ﬂow of recommendation (Figure 1) begins with a

user connecting to the application, using his/her

credentials from the music platform and later the

social network credentials.

Next, the user deﬁnes a new group or chooses one

of the already existing ones, from which he/she will

listen to songs (activities that will be done in a group).

The application then retrieves data from the music

activities such as the name of singers, or the bands,

the name of songs, and the total number of times the

user listened to the songs. In addition, the application

also retrieves data from the social activity of the group

of friends, including the user’s own, such as mentions,

comments, likes, and information sharing.

The hypothesis is the friends in the group

already acquired items in e-commerce. Therefore,

the preference data of friends are known in the

e-commerce. Thus, in the controlled experiment, the

group preferences in the musical platform are also

collected.

Using MF, more precisely the ALS, the data

of the musical activities are used to generate

recommendations of music for each member in the

group.

Figure 2: Social Tracks’ recommendation process.

Social data is used to calculate Leader and

Similarity Factors. The Expert Factor, as opposed to

the others, is calculated with the data of the musical

activity of each user. Since all these factors have been

computed, the inﬂuence of each individual in relation

to the group is calculated and also the individual

recommendation. Thus, the system aggregates

the individual recommendations considering the

inﬂuence, obtaining the recommendation of the set

Social Tracks: Recommender System for Multiple Individuals using Social Inﬂuence

367

of songs for the initially selected group. At the

end of the process, users receive, through the mobile

application, the set of recommended songs, which

can be heard and rated. The rating makes possible

to evaluate whether the result of the recommendation

was assertive or not for the group.

The individual recommendations calculated using

ALS is the most expensive step in the ﬂow processing

(Figure 2) because the set of data is very large since

the database has thousands of users and thousands of

songs to determine the latent factors.

Although the calculation of the inﬂuence factors

algorithm (Figure 2) is computationally expensive, it

is done in a short time, since the groups are limited

to a maximum of 5 individuals. This limitation was

imposed by the computational resources employed

in the experiment to meet the system requirements

making recommendations in an acceptable time.

4.2 Controlled Environment

The controlled experiment use a social network

(SocialTracks Mastodon) created to ensure that

information exchange occurs in the social network as

well as music information in the music platform for

the group that will receive the recommendation.

Mastodon (https://mastodon.social) was used to

create a social network; it was chosen because it is

open source, easy to use and it has features similar to

Twitter, which is a widely used social network with

fewer restrictions on accessing its users’ data.

The instance of Mastodon was created to hold up

to 100 users. All the computational limitations were

dictated by the existing resources.

In the second stage of the research the experiment

will be repeated in an uncontrolled environment to

compare the results. The result of the second stage of

the research will be timely disclosed in a future paper.

4.3 Data for the RS Training

4.3.1 Music Service Platform Dataset

Last.fm’s public API was used to collect the dataset

that served as the basis for training the ALS

algorithm. Last.fm API is an online aggregator of

musical data that allowed the crawling of 105,655

songs, rated by 41,242 users, with a total of 309,986

implicit ratings. In this API, all users’ data are

publicly available. This dataset was used to simulate a

music service platform with a considerable number of

users and enough music to generate recommendations

through the ALS algorithm. Table 1 details this

dataset. The User-Song data in this table is an

example of implicit rating because it provides no

direct rating of songs, but rather the number of times

a user has heard a particular song.

Table 1: Last.fm dataset information details.

Data type Quantity Details

Song 105.655 Name, duration

and artist (name

of singer, band or

group)

User 41.262 Name, country, age,

and playcount (total

number of times

the user listened to

songs)

User-Song 309.986 User, song and

playcount (number

of times the user

listened to a

particular song)

4.3.2 Individual Musical Preferences Dataset

Spotify was the music platform used to obtain

individual musical preferences of the experiment’

participants (Figures 1 and 2). It was chosen because,

besides being a widely known platform, it provides an

easy-to-use REST API, allowing an easy integration

with web/mobile application. Spotify API returns

JSON metadata about artists, albums, and music

tracks that the logged-in user listened to in Spotify.

The preferences of each participant are collected

when he/she logs into the application. Then, these

data are preprocessed and added to Last.fm’s dataset

to carry out the training. This was necessary because

of the ALS algorithm, which only generates the

recommendation of new items if the user is part of

the training dataset. Part of the preprocessing step

was ﬁltering individual musical preferences to contain

only the songs matching the set of songs available in

the Last.fm dataset.

The algorithm then evaluates the users’

recommendations based on the latent factors. A

list of the 50 best-rated songs for each user is

generated.

5 TEST AND RESULTS

The results presented here constitute a proof of

concept for the proposed RS. For this, 11 participants

allowed forming 10 groups with a varied number of

members.

KDIR 2019 - 11th International Conference on Knowledge Discovery and Information Retrieval

368

Table 2: Computation of inﬂuence factors.

Group UserId F

exp

lead

sim

in f

2 0.8184 0.2619 0.1666 0.4156

11 0.6260 0.3611 0.0416 0.3429

5 0.1909 0.4166 0.0833 0.2303

8 0.3644 0.2936 0.0416 0.2332

9 0.4945 0.4448 0.4771 0.4722

6 0.4016 0.2758 0.1953 0.2909

3 0.6037 0.2792 0.5225 0.4685

9 0.5660 0.5555 0.5 0.5405

6 0.4339 0.1111 0.5 0.3483

6 0.5046 0.4652 0.0750 0.3483

11 0.6319 0.1583 0 0.2634

9 0.5805 0.5069 0.1583 0.4152

4 0.5176 0 0 0.1725

8 0.2651 0.4527 0.0833 0.2670

2 0.7021 0.4166 0.3571 0.4919

4 0.3179 0.2444 0.4625 0.3416

3 0.4798 0.3388 0.3196 0.3794

9 0.7668 0.2500 0.25 0.4222

5 0.2273 0 0 0.0757

8 0.5058 0.5833 0.25 0.4463

2 0.6583 0.3737 0.5 0.5106

3 0.3416 0.2929 0.5 0.3781

6 0.6114 0.5256 0.1369 0.4246

10 0.0903 0 0.0625 0.0509

5 0.2196 0.2896 0.0333 0.1809

7 0.7853 0.3169 0.0750 0.3924

3 0.7098 0.5343 0.1410 0.4617

2 0.9887 0.4166 0.5 0.6351

10 0.0112 0.2500 0.5 0.2537

2 0.7381 0.4470 0.1450 0.4434

3 0.6078 0.4789 0.1290 0.4052

10 0.0562 0.3161 0.0357 0.1360

9 0.5493 0.5258 0.1254 0.4002

8 0.3573 0.2825 0.0177 0.2192

1 0.7486 0 0 0.2495

7 0.6807 0.3371 0.0274 0.3484

4 0.5178 0.1600 0.0542 0.2440

6 0.5006 0.4659 0.0997 0.3554

5 0.2060 0.2825 0.0187 0.1691

Table 2 exhibits the inﬂuence factors calculated

for each member from each group.

Table 2 shows each group and its members. User

data has been masked for privacy reasons, being the

users identiﬁed by an ID varying from 1 to 11 (ﬁeld

userId).

To generate the recommendation for a group, the

following is computed:

• individual recommendations made (Figure 2)

using user’s data retrieved from Spotify, and

• user’s inﬂuence factor calculated using data

retrieved from a social network.

The similarity factor (F

sim

) is inﬂuenced by the

user’s interaction in the social network; in other

words, the F

sim

depends on how active a user is in the

social network. This factor can be different for two

users, even though they are very similar. For example,

consider two users i and j that give many likes to

posts made by each of them, mutually. However,

user i usually shares more posts from user j than the

opposite; then F

sim

for user i will be greater than for

user j. Besides, F

sim

for a user may be zero if he/she

only usually posts in the social networks, without

sharing, or giving likes, to the posts from the other

friends. In the same way, the leader factor (Flead)

might be zero for an individual that did not receive

likes or mentions from the others in the group.

The ﬁnal tests with an appropriate number of users

(Krejcie and Morgan, 1970) will be performed and

presented in a future paper.

5.1 Group Recommendation Results

The Social Tracks Application (Figure 1) is also

employed to allow users to evaluate the result of

recommendations. For this, 20 tracks are returned

to each participant in the group that gives a grade

from 1 to 5 to each music. The root mean square

error (RMSE) (Gonzalez-Camacho and Alves-Souza,

2018) is used to measure the difference between rates

given by users and the system. The GRS generates

values in the range of [0, 2] (Eq. 10). Therefore, GRS

results were normalized to values in the interval of [1,

5] (Eq. 11).

System

(g, j) =



g, j

max(RG

g, j

)

∗ 4



+ 1 (11)

Where R

System

(g, j) is the normalization of j’s

item rating for a given g group. RG

g, j

is deﬁned in

Eq. 10.

Tables 3 and 4 introduce part of the

results for 2 groups. All the results

from the experiment are available in

https://github.com/SocialTracksDataAnalise.

Table 3 shows three tracks recommended to group

1, and Table 4 shows those to the group 5. Field RG

g, j

illustrates the prediction values given by the system

and R

system

is its respective normalized value. UserId

identiﬁes each member in the group; R

is the track

rating made by a respective member and R

is the

average of the ratings for the track, calculated from

, for the group.

Social Tracks: Recommender System for Multiple Individuals using Social Inﬂuence

369

Table 3: Part of track rating results of group 1.

Group Track

Name

g, j

system

UserId R

Sex on

Fire

1.62 4

8 5

5 1

2 5

11 5

The

Scientist

1.66 4

8 4

4.25

5 5

2 4

11 4

the End

1.60 4

8 4

5 5

2 3

11 4

Table 4: Part of track rating results of group 5.

Group Track

Name

g, j

system

UserId R

Sex on

Fire

1.35 3

2 5

4.333 4

4 4

The

Scientist

1.42 4

2 4

3.663 3

4 4

Mr.

Brightside

1.43 4

2 2

3.333 3

4 5

RMSE is given by:

RMSE =

∑

i=1

∑

j=1

(

i, j

− R

i, j

)

(12)

K =

∑

i=1

∑

j=1

1 (13)

In eq.12,

i, j

= R

system

(g, j), it is the rating that

the system made for a track j, which is recommended

for group i. R

i, j

is the average rating that a group

i made for a recommended track j. N is the total

number of groups participating and M

is the number

of recommendations made for the group i,

The RMSE value varies between 0 and 4

(Contratres et al., 2018), because as



i, j

− R

i, j



≤ 4,

we have 0 ≤ (

i, j

− R

i, j

)

≤ 4

, then:

0 ≤

∑

i=1

∑

j=1

(

i, j

− R

i, j

)

≤ 4 (14)

The RMSE value obtained for this experiment

was:

RMSE = 1.4554 (15)

According the RMSE value, the variation between the

ratings given by GRS and members of the groups was

1.5 approximately.

Figure 3 exhibits two distribution of ratings

of the pieces of music, one given by the system

−system

) and the other, given by the group of

users (R

−av

). These results conﬁrm that the music

recommendations made for the groups were well

appreciated, once the ratings made by the groups were

1.45 points above the prediction.

Figure 3: Distribution of prediction given by GRS and by

the groups.

6 CONCLUSIONS

The proposed RS makes the recommendation for

groups of individuals considering the inﬂuence factor

among them, which is calculated based on three

factors: expert, leader and similarity. In our proposal,

these factors are calculated using information from

a social network. Controlled experiment allowed

testing our proposal, guaranteeing that the members

of the groups had some activity in the social network.

For a proof of concept, the controlled experiment

was conducted with 10 users to demonstrate the

effectiveness of our proposal. As a result, the

assertiveness of the RS, computed based on RMSE

was 68.7%. As future work, we intend to integrate

the inﬂuence factor within the ALS algorithm

for calculating the latent factors. As lessons

learned, besides the difﬁculty in managing users

for taking part in the experiment, a user-test based

on user-experience should be conducted to help to

prepare the user instruction material to facilitate

her/his participation. The contributions of this

paper are: (i) the inﬂuence factor that consider the

capacity of leadership, expertise in a subject and the

preference similarity among individuals that will do a

common activity. These factors are calculated using

only information retrieved from social networks;

(ii) the use data from social network to improve

recommendation. As a future work the inﬂuence

factor, also with data extracted from social networks,

KDIR 2019 - 11th International Conference on Knowledge Discovery and Information Retrieval

370

will be proposed to improve the recommendation for

a cold-start scenario.

ACKNOWLEDGEMENTS

The authors are grateful for the support given

by S

ao Paulo Research Foundation (FAPESP).

Grant #2014/04851-8, and the support given by

Ita

u Unibanco S.A. trough the Ita

u Scholarship

Program, at the Centro de Ci

encia de Dados (C

D),

Universidade de S

ao Paulo, Brazil.

REFERENCES

Al-Hassan, M., Lu, H., and Lu, J. (2015). A

semantic enhanced hybrid recommendation approach:

A case study of e-Government tourism service

recommendation system. Decis. Support Syst.,

72:97–109.

Contratres, F. G., Alves-Souza, S. N., Filgueira, L.

V. L., and DeSouza, L. S. (2018). Sentiment

analysis of social network data for cold-start relief

in recommender systems. In New Contributions in

Information Systems and Technologies. Advances in

Intelligent Systems and Computing. in press.

Deng, S., Huang, L., and Xu, G. (2014). Social

network-based service recommendation with trust

enhancement. Expert Systems with Applications,

41(18):8075–8084.

Desrosiers, C. and Karypis, G. (2011). A Comprehensive

Survey of Neighborhood-based Recommendation

Methods, chapter 4. Springer US.

Gonzalez-Camacho, L. A. and Alves-Souza, S. N.

(2018). Social network data to alleviate

cold-start in recommender system: A systematic

review. Information Processing & Management,

54(4):529–544.

Guo, J., Zhu, Y., Li, A., Wang, Q., and Han, W. (2016). A

Social Inﬂuence Approach for Group User Modeling

in Group Recommendation Systems. IEEE Intelligent

Systems, 31(5):40–48.

Hu, Y., Volinsky, C., and Koren, Y. (2008). Collaborative

ﬁltering for implicit feedback datasets. Proceedings

- IEEE International Conference on Data Mining,

ICDM.

Koren, Y. and Bell, R. (2011). Advances in Collaborative

Filtering, chapter 5. Springer US.

Koren, Y., Bell, R., and Volinsky, C. (2009). Matrix

factorization techniques for recommender systems.

Computer, 42(8):30–37.

Krejcie, R. V. and Morgan, D. W. (1970). Determining

Sample Size for Research Activities . Educational and

Psychological Measurement, 30(3).

Lalwani, D., Somayajulu, D. V. L. N., and Krishna, P. R.

(2015). A community driven social recommendation

system. Proc. - 2015 IEEE Int. Conf. Big Data, IEEE

Big Data 2015, pages 821–826.

Lian, D., Ge, Y., Zhang, F., Yuan, N. J., Xie, X., Zhou,

T., and Rui, Y. (2016). Content-aware collaborative

ﬁltering for location recommendation based on human

mobility data. Proc. - IEEE Int. Conf. Data Mining,

ICDM, 2016-Janua:261–270.

Masthoff, J. (2011). Group Recommender Systems:

Combining Individual Models, chapter 21. Springer

US.

Masthoff, J. and Gatt, A. (2006). In pursuit of satisfaction

and the prevention of embarrassment: affective state

in group recommender systems. User Modeling and

User-Adapted Interaction, 16(3-4):281–319.

Prando, A., Contratres, F., Alves-Souza, S., and deSouza,

L. (2017). Content-based recommender system using

social networks for cold-start users. In Proceedings of

the 9th International Joint Conference on Knowledge

Discovery,Knowledge Engineering and Knowledge

Management. in press.

Quirino, G. Z., Mals, N. P., Groterhorst, V. M., De Souza,

S. N., and De Souza, L. S. (2015). Meneduca

- Social school network to support the educational

environment. In Proceedings - 2015 41st Latin

American Computing Conference, CLEI 2015.

Resnick, P. and Varian, H. R. (1997). Recommender

systems. Communications of the ACM, 40(3):56–58.

Ricci, F., Rokach, L., and Shapira, B. (2011). Recommender

Systems Handbook. Springer US.

Wang, Z. and Lu, H. (2014). Online recommender

system based on social network regularization.

8834:487–494.

Zhang, Y., Chen, W., and Yin, Z. (2013). Collaborative

ﬁltering with social regularization for TV program

recommendation. Knowledge-Based Syst.,

54:310–317.

Zhao, W. X., Li, S., He, Y., Chang, E. Y., Wen, J. R.,

and Li, X. (2016). Connecting Social Media to

E-Commerce: Cold-Start Product Recommendation

Using Microblogging Information. IEEE Trans.

Knowl. Data Eng., 28:1147–1159.

Zhu, H. and Huberman, B. A. (2014). To Switch or

Not To Switch: Understanding Social Inﬂuence in

Online Choices. American Behavioral Scientist,

58(10):1329–1344.

Social Tracks: Recommender System for Multiple Individuals using Social Inﬂuence

371