POI-based Recommender System for the Support of Academics in a

Smart Campus

Guilherme Xavier de Carvalho

1 a

, Isabela Gasparini

1 b

, Guilherme Medeiros Machado

2 c

Leandro Krug Wives

2 d

and Jos

e Palazzo M. de Oliveira

2 e

Santa Catarina State University (UDESC), Joinville, SC, Brazil

Informatic Institute - PPGCC, Federal University of RS (UFRGS), Porto Alegre, Brazil

Keywords:

Recommender Systems, Smart Campus, Points of Interest.

Abstract:

Recommender Systems are software used to suggest user items in a personalized and automated way. When

combined with Points of Interest (POI), they can set locations as referable items. This type of approach is

useful when the amount of POI available for the user is large. In the context of Universities, students have

different needs and have to look for different locations to experience the Universities’ resources. The goal

of this paper is to present a POI-based Recommender System to improve student’s well-being and to support

their academic journey in a Smart Campus. The recommender system was implemented by an application

called AONDE, which was used by 110 students, where 63 answered a satisfaction questionnaire allowing the

data collection needed for the the system evaluation. An accuracy of 61% in the recommendations of items to

students was measured, as well as a high satisfaction rate, where 90.5% of respondents said they were satisﬁed

or very satisﬁed with the locations suggested by the app. The purpose of this experience paper is demonstrate

that the approach here described proved to be useful for students’ routine, impacting positively their academic

journey.

1 INTRODUCTION

Recommender systems (RS) were ﬁrst proposed as a

solution to deal with the problem of user cognitive

overload, where the amount of information to be an-

alyzed exceeds the user capability (Machado et al.,

2018). RS is a subclass of information ﬁltering sys-

tems that seeks to predict the “rating” or “preference”

a user would give to an item (Ricci et al., 2011). They

are used in a variety of applications, such as sugges-

tion of movies (e.g., Netﬂix), music (e.g., Spotify), or

videos (e.g., YouTube). When combined with Points

of Interest (POI), they can set locations as referable

items. This type of approach is useful when the

amount of POI available for the user is bigger than

the user capacity to analyze the full set. Therefore,

there is a need to customize recommendations to meet

the interests and needs of the user. Such systems can

https://orcid.org/0000-0002-9203-662X

https://orcid.org/0000-0002-8094-9261

https://orcid.org/0000-0001-5283-9228

https://orcid.org/0000-0002-8391-446X

https://orcid.org/0000-0002-9166-8801

be used in intelligent scenarios to improve the quality

of life of individuals in this environment; since data

ﬁltering reduces the number of items associated with

users’ domain, their cognitive and informational load

are reduced.

The goal of this paper is to present a POI-based

Recommender System in a Smart Campus to sup-

port the academic journey in the environment. For

the system development, a literature review was per-

formed in search of related work and the basic con-

cepts were identiﬁed. The recommender system was

created, aligned primarily with the interests and needs

of its’ target audience. The algorithm implemented

represents the users and the items to be recommended

through a vector of tags, which are keywords that de-

scribe such an entity. Both users and items are struc-

tured in matrices and weighted by the metric of Term

Frequency–Inverse Document Frequency (TF-IDF).

Finally, the similarity between items and users is cal-

culated and generates the recommendations that are

presented through a web application, called AONDE,

which allows interaction with the user.

AONDE was used by 110 students of the Santa

Catarina State University - UDESC, and 63 answered

398

Xavier de Carvalho, G., Gasparini, I., Machado, G., Wives, L. and M. de Oliveira, J.

POI-based Recommender System for the Support of Academics in a Smart Campus.

DOI: 10.5220/0009514003980405

In Proceedings of the 12th International Conference on Computer Supported Education (CSEDU 2020) - Volume 2, pages 398-405

ISBN: 978-989-758-417-6

a satisfaction questionnaire and thus allowed the col-

lection of inputs used in the evaluation of the system.

Through these, an accuracy of 61% in the recommen-

dations of items to users was measured, as well as

a high satisfaction rate, where 90.5% of respondents

said they were satisﬁed or very satisﬁed with the lo-

cations suggested by the app.

2 DEVELOPED RECOMMENDER

SYSTEM

In this section, we present the system that was de-

signed and developed to provide POI recommenda-

tions inside a smart campus. The strategy used in the

system is content-based since we did not have an an-

notated dataset (items previously evaluated). The sys-

tem is developed to attend two speciﬁc student stereo-

types, i.e., newcomers and veterans. More speciﬁ-

cally, it is designed to help newcomers, i.e., students

that recently arrived at the campus and do not know

the environment, the routines, neither the resources

available to them. It is also able to help veteran stu-

dents to get involved with current activities and news,

since we found that they are usually more focused on

ﬁnishing their courses and do not have time to keep

up to date.

Based on these stereotypes, and also the user inter-

action, the system is able to identify the current user

needs and interests, classifying and prioritizing POI

accordingly to each user proﬁle. Thus, not only the

most interesting items are recommended, but also the

ones most necessary to the academic life of each pro-

ﬁle.

2.1 POI and User Modeling and

Representation

Both items to be recommended by the Recommender

System and user interests are represented within the

application through tags, keywords that deﬁne an el-

ement in a short, straightforward and clear manner.

Therefore, it was necessary to deﬁne the best tags

for each POI. A questionnaire was designed to col-

lect students’ perceptions of the locations found on

the university campus, as well as how they would de-

scribe them using tags.

The questionnaire was organized into three sec-

tions, which sought to generate mechanisms so that

we could understand, through user responses, some

aspects such as: which locations on campus were in-

teresting for students to know, with which tags certain

locations could be described, what was the proﬁle of

freshman and senior student, what were the interests

of students as members of a college campus, what was

the relationship of the student’s course and the places

of interest.

Before being applied to the general public, a pilot

test was accomplished with students of the Human-

Computer Interaction discipline. This pilot test was

conducted to verify if the questionnaire was well

structured and organized and the interaction to an-

swer were analyzed. Such students contributed to the

reﬁnement of the research instrument and once val-

idated, it was sent for access to all students on the

university through institutional emails and shares in

social networking groups of the institution. The ques-

tionnaire was available from August 6 to September

12, 2019; and 133 responses were collected. It is note-

worthy that there was at least one student response

from each of the 21 courses offered by the university,

among undergraduate and graduate students.

Based on information from this questionnaire

and three standard university documents the Fresh-

man Manual, the Extension Project Guide, and the

Pedagogical Curriculum Guidelines from UDESC,

we were able to deﬁne the representation of POIs,

courses, user interests and freshman stereotype using

a folksonomy.

According to Pereira e Silva (Pereira and da Silva,

2008) the folksonomy technique represents an initia-

tive to assist in the process of organization and at-

tribution of meaning to the contents available on the

web, as well as for the treatment of information over-

load. Thus, it is a content analysis and organization

approach based on the view of the interviewed audi-

ence and the author. The process of categorization of

information is based on the three pivots of folkson-

omy, the user - who performs the categorization; the

object - which is categorized; and the tags - which cat-

egorize this object. It is worth mentioning that this is

exactly the path to be followed through the question-

naire directed at the students.

Applying folksonomy can be understood as two

very distinct moments: information categorization

and information retrieval. In the ﬁrst, users are given

a freedom of work, which enables them to generate

hypotheses and great information through a practical

and inexpensive process of both time and cognitive

effort. The second moment is related to the classiﬁca-

tion of terms chosen during categorization, in short, is

to evaluate the quality of the answers and make deci-

sions from them. At this time freedom of work is on

the side of the experimenter, so it is possible to make

decisions by looking at the inputs collected from users

in the ﬁrst phase of the approach and also from the

convictions that the project scope already has (Pereira

POI-based Recommender System for the Support of Academics in a Smart Campus

399

and da Silva, 2008).

The following sections explain how each of these

elements was elaborated, considering the glued sam-

pling and the applied folksonomy techniques, to be

implemented in the web application.

A survey had been carried out to deﬁne primarily

which POIs would be recommended by the system.

After data collection, it was possible to understand

which of those POIs were indeed relevant to the stu-

dents’ context. Such deﬁnitions emerged by applying

folksonomy techniques to the information base col-

lected and thus reached the number of 86 recommend-

able POIs within the UDESC Smart Campus.

Once the POIs were chosen, attention turned to

deﬁning the tags that would describe a particular loca-

tion. At this time, the students’ answers were also an-

alyzed in order to understand their perceptions about

the content that would serve as a description of a par-

ticular POI. It was also through them that it was pos-

sible to extract the keywords for each of the recom-

mended places. However, it is important to note that

not all tags that make up the RS came directly from

the students’ verbalization or responses in the ques-

tionnaire.

Having the folksonomy presented by Pereira and

Silva (2008) as the basis for this process of analysis

and deﬁnition, decisions were made looking at data

sampling based on the scope of the project. Thus,

some tags were elaborated through the interpretation

of the author of this work under the proﬁle of the stu-

dents who participated in the research, relating the

academic journey of a student with the locations listed

as POI. We deﬁned 83 distinct keywords, which were

scattered among the items to be recommended con-

textually, where each POI has at least 3 tags describ-

ing it.

Still on this process it is useful to emphasize that

the task of establishing the keywords for the POI

needs some care. Since if many POIs have a high tag

similarity rate, it will be difﬁcult for the RS to differ-

entiate the specialties of each recommendation item.

Therefore, such a system may recommend items that

are not necessarily interesting to a user but which have

a high similarity index between the suggested tags.

On the other hand, if each POI has tags that are sig-

niﬁcantly speciﬁc to it, it will be difﬁcult to correlate

the items to be recommended with the user’s interests

more generally, thus requiring explicit and individu-

alized statements about the interest in each POI.

We have designed two scenarios to show the fea-

sibility of the system, one for each stereotype previ-

ously deﬁned.

The ﬁrst one considers a student named Aurora,

a freshman student at the university. She is informed

by the administration that a application is available

to help students get localized and to know the uni-

versity. She installs it on her smartphone and ﬁlls

in her proﬁle and chooses some categories of inter-

est: Culture, entertainment and research. Right af-

ter, the system starts calculating the recommendations

and shows different POIs. Aurora starts her tour in

the campus, very excited, and starts sharing her expe-

rience with colleagues, which, in turn, start using the

application too. She ﬁnds out that some recommen-

dations, even if her colleagues have chosen the same

categories, are different (most of them are older than

her in the university). Hers interests are more related

to bureaucratic activities, needed in her initial stage.

The second time she used the application, the proﬁle

was already ﬁlled and the recommendations were im-

mediate, but she was able to change her interests, per-

sonalizing the recommendations.

The second scenario considers a student named

Tobias, who is in the 3

year. He checked the bulletin

board and found the information about a new app.

Surprised by such news, he ﬁrst wondered “When did

they create it?”. Even suspicious about what it was re-

ally about, he accessed the application. Tobias created

his proﬁle by entering his personal details and select-

ing the categories that seemed convenient. A few mo-

ments later the system presented him with places that

would be of interest to him, respecting the newly cre-

ated proﬁle. Analyzing the suggestions that appeared

on his list, Tobias realized that even though he was

a veteran, there were still several locations on cam-

pus he didn’t know about. The student was interested

in the functionality of the application, because it be-

came more updated with the actions that take place at

the university. Using other times, he noted that he can

modify his proﬁle and disregard elements that are as-

sociated with activities he already knows, such as the

restaurant for example.

As our RS is content-based, the way in which the

content of the items is represented should be made

explicit. In this case, the items to be recommended

through the system are the POI. Therefore, these are

modeled through a structure where each POI is asso-

ciated with a set of keywords that describe that re-

spective POI, functioning as tags. An array of char-

acteristics (items x tags), where each line represents a

POI, so there will be a vector for each POI.

Considering these aspects and the feasibility of us-

ing tags to describe items, the second phase of the

content representation process deals with the weight-

ing of the items. According to Jannach et al. (Jannach

et al., 2010), the purpose of a content-based recom-

mender approach is not to maintain a list of items and

their static meta-information, but to categorize their

CSEDU 2020 - 12th International Conference on Computer Supported Education

400

relevant keywords in the form of weights. Therefore,

items are typically described using the TF-IDF en-

coding format (Term Frequency – Inverse Document

Frequency) (Cantador et al., 2010) . Following the

approach via tags, where tags are used instead of doc-

uments, the TF-IDF encoding was adjusted and ac-

cording to the literature it can be applied in two ways:

in the ﬁrst, only the Term Frequency (TF) is calcu-

lated, in the second the full scope of the formula is

taken into account. In this work, only the ﬁrst version

of the formula is applied. Which describes an item

by means of tags, and precisely by using this artiﬁce

there is no need to check the inverse frequency of a

term in the document (Cantador et al., 2010), after all

there are only a set of tags describing each item to be

recommended. Thus, the proﬁle of an item is repre-

sented by the vector i

= (i

n,1

,...i

n,l

) expressed by the

equation below:

n,l

= t f

) (1)

Where t f

) corresponds to the number of times

(t f ) an item i

was highlight with the tag t

to a certain

item i

n,l

At the end of the weighting process, the set of

items will be represented as a vector of the weights

computed by the TF.

The representation of the user occurs in a similar

way to that of the item. In this scheme, user proﬁle

is represented through a matrix of interest structure

(user x tags), where each line represents a user, so

there is a vector for each user. Each cell in the ma-

trix is linked to a speciﬁc tag and through this feature

it is possible to know which tags a user is associated

with. It is worth mentioning that these tags must be

collected through the user’s past interactions or ex-

plicitly asking the user when the user creates a proﬁle

in the application.

Weights are also plotted in a vector space, adjusted

noting only by interest, but also by the user’s needs.

Therefore, the weighting of a user’s proﬁle consists of

the vector u

= (u

m,1

,...u

m,l

), expressed by:

m,l

= t f

) (2)

Where t f

) corresponds to the number of times

(t f ) that user u

has been associated with tag t

for

a given user u

m,l

. After calculating the TF the user

vector proceeds to another weighting step, since still

in the user representation, it is necessary to consider

the recommendations for the freshman user proﬁle.

For this purpose, a calibration in the value of the items

that make up the vector of interest of this type of user

is incorporated.

Knowing that the user is a freshman, it is pos-

sible to evaluate the interest vector of that user to

give more or less weight to certain interests, through

pre-processing.The evaluation in this stage considers

the stereotype of a freshman user. This stereotype is

a user vector F, which followed the representation

established for the user, but which already had pre-

deﬁned tags according to the items (POI) mapped as

necessary for a freshman.

These POI considered necessary for freshmen

were taken from the University Freshman Manual,

which deﬁnes some main locations within the Univer-

sity campus. Therefore, this F stereotype is applied to

all users who use RS and are freshmen, that is, they

have been in the university for less than a year. It

is, therefore, a calibration imposed on the vector al-

ready calculated for these users. Such calibration is

performed by means of a simple average between the

vector of the user already weighted and the vector F,

in order to increase the weight of tags related to the

items deﬁned as necessary for the freshmen. After

the user modeling process is ﬁnished, it is possible to

proceed to the recommendation algorithm, where the

similarity between item and user is calculated.

2.2 Finding Relevant POI

After the item and user were processed, the third stage

of processing deals with the similarity between the

item and user proﬁles. This is when the recommen-

dation is actually generated. For that, the cosine sim-

ilarity method, established as state of the art in this

context is applied (Ricci et al., 2015).

The proximity calculation can be performed using

the following equation, which is applied to the item

and users weighted vectors only by TF:

simCos

t f

) =

∑

k=1

t f

) · t f

)

∑

k=1

(t f

))

∑

k=1

(t f

))

(3)

Where simCos

t f

) corresponds to the simi-

larity coefﬁcient between the user vectors and items

(value belonging to the set [0.1]); t f

) is the num-

ber of times that the user u

has been associated with

textit tag t

; and t f

) deals with the number of times

that the item i

has been marked with the textit tag t

It is worth mentioning that for the calculation

of similarity, the corresponding mathematical expres-

sion was implemented in an algorithm in the Python

language, without the use of any speciﬁc library.

At the end of this step, the items to be recom-

mended to the user have already been deﬁned. There-

fore, they are subject to suggestion and can be pre-

sented to the user.

POI-based Recommender System for the Support of Academics in a Smart Campus

401

2.3 Architecture of the System

With regard to architecture, the system is based on

the client-servers concept. A structure where service

providers, i.e. servers, are separated from those who

request the service, e.g. customers (Lee et al., 2004).

This relationship is conducted in such a way that cus-

tomers use server resources, without having direct ac-

cess to them. For that, it is on the basis of the ex-

change of requests that the application works.

The architecture is organized in three layers: pre-

sentation layer, business layer and data layer. This

model is implemented to organize the main entities

that make up the app. The presentation or interface

layer is the layer that interacts directly with the user.

Therefore, it is through it that requests for consulta-

tions are made and the results are displayed. The busi-

ness layer, on the other hand, is the logical part, where

the RS itself is, in order to relate to the other two lay-

ers for its full functioning. The third layer of this orga-

nization is the data layer. It is the entity composed of

the information repository that receives requests from

the business layer and executes these requests in the

database.

The user receives recommendations through an

application. When accessing it, the user is interacting

in the presentation layer, without the need to know the

processes behind the RS itself. It is worth mentioning

that the application is on the client side, in the client-

server relationship, and relates to the server side when

sending and receiving data.

On the server side are the data and business layers.

Respectively, the ﬁrst layer represents the database

with all the necessary information in the recommen-

dation process that is manipulated. The second, on the

other hand, stores the recommender’s business rules,

which is in fact the logical and coded deﬁnitions of

the RS. Once the architectural organization is estab-

lished, it is possible to specify the technical aspects

of the system in each of the layers. For the presen-

tation layer, a Web App was developed, a responsive

web application accessed from any browser installed

on computers or smartphones. Such a system was de-

veloped using HTML5, CSS3, JavaScript and PHP

technologies.

The Web App communicates with the other layers

through RESTful API services developed in Python.

It is worth mentioning that the application design

process was guided by the Double Diamond model,

based on the Design Thinking approach. The data

layer of the system is structured by the PostgreSQL

DBMS.

As for the business layer, this is where the logical

part of the recommendation algorithm is located. RS

was implemented in Python and using the tools of the

language itself, a server was also built. Such business

layer resources were hosted in a structure that allows

users to access remotely. In this sense, the hosting

was performed on a server within the University’s in-

frastructure. Figures 1 and 2 present the recommen-

dation from AONDE application.

Figure 1: Recommendation of AONDE system in a desktop.

Figure 2: Recommendation of AONDE system in a smart-

phone.

For each recommended POI, the student can give

his/her feedback, selecting the “like” or “dislike” but-

tons presented just below the description of each POI.

3 EVALUATION AND RESULTS

The AONDE application was available to the Uni-

versity students and the experiment was carried out

between October 21 and 31, 2019. The release was

communicated through an institutional email, which

CSEDU 2020 - 12th International Conference on Computer Supported Education

402

would make it possible to reach a large part of the

students on the university campus.

During this current period of the experiment, 110

students registered in the application.The inputs gen-

erated from the interactions where they were mon-

itored periodically through the evaluation question-

naire and the database. After this period the access

to the application has been restricted and the data ma-

nipulation and analysis phase has started.

For the process of evaluating the AONDE appli-

cation and its recommender system, the metrics of

accuracy of the Recommender System (RS) and user

satisfaction regarding the use of the application were

established. So that such metrics could be analyzed,

two instruments were created. The ﬁrst is related to

the formalism of precision and recall used to calcu-

late the accuracy of an RS, which is combined with

a feedback system (like and dislike) inserted in each

recommended POI. While the second is a question-

naire created to collect evidence related to the usabil-

ity of the system and user satisfaction.

The analysis of the data from this experimental

mechanism is based on the calculation of precision

and recall. In order to precision and recall to be calcu-

lated in a Recommender System, the top − k items of

greatest relevance delivered by the system to the user

are used. In the context of our system, it was deﬁned

that the 8 items with the highest recommended simi-

larity rate would represent this set. Precisely because

this is the number of POIs shown on the ﬁrst page.

Therefore, the accuracy and recall were calculated for

each user who interacted with the system.

As an example, imagine that a student named Bob

accessed the system and received his recommenda-

tions after creating his proﬁle. Bob interacts with the

system, navigating between pages, give “likes” and

“dislikes” in some of the suggested locations. This

interaction provides the necessary data to assess the

accuracy of the recommendations generated for Bob.

Suppose he liked 5 of the ﬁrst 8 POIs presented, that

is, those on the ﬁrst page, and that in total he liked

13 of all 86 recommended items, including those on

other pages. In this case, the precision would be

5/8 ≈ 0.63, that is, 63%, where 5 corresponds to the

number of items deﬁned as relevant selected by the

user, and 8 the total number of items considered rele-

vant.In this case the recall would result in the propor-

tion 5/13 ≈ 0.38, that is, 38%, where 5 corresponds

to the number of items deﬁned as relevant selected by

the user, and 13 the total number of items selected by

the user.

Applying this logic and analyzing the data col-

lected through the interactions of users with the items

recommended by AONDE, we can measure an accu-

racy of 61% and a recall equal to 66%, this having

limited to 8 the number of items considered relevant.

An accuracy of 61% represents that on average, for

every 10 items recommended, 6 met the expectations

of users, were in line with the proﬁle and aligned with

the context. Therefore, it can be said that 61% of the

suggestions through AONDE were in fact assertive.

Thus, the probability that a selected item is the one

considered relevant for a student who interacted with

AONDE is 61%.

The RS recall was 66%, this represents the aver-

age number of times a relevant item is selected from

the total number of items selected. Thus, the greater

the recall, the greater the number of items considered

relevant to have been selected from the total. In this

context, with a 66% recall, it can be said that of all

items selected by the student during use, 66% were

among those considered relevant (top-8).

In short, the metrics collected through the data

samples point to a system with good recommendation

accuracy, considering the values of 61% for precision

and 66% for recall. The numbers indicate that in at

least more than 60% of the recommendations gener-

ated by the RS were relevant.

Of the 110 students who used AONDE applica-

tion and received recommendations, 63 students (i.e.

57%) answered the satisfaction questionnaire. Of

these students, 52.4% are male and 49.2% are fresh-

men. Students from 16 different courses participated

in the assessment (from 21 courses available in the

campus).

Regarding some usability issues from the ques-

tionnaire, they were presented in a 5-point likert scale.

Figure 3 presents the results. As Figure shows, most

students strongly agreed with the usability issues of

the app.

Regarding the question in which the student is in-

vited to say how useful for the academic journey of

a university student is to know the Points of Interest

(POI) recommended by the system, using a scale from

0 to 10, where 0 represents a lot useless and 10 very

useful, the Figure 4 presents the results. It can be seen

that the vast majority of academics agree that know-

ing such places is of great use, when voting 41.3% for

option 9 and 38.1% for option 10, totaling 50 of the

63 users who participated in the evaluation.

Students were also asked how satisﬁed they were

with the recommendations they received. As can be

seen in the Figure 5, the vast majority of those who

used the system and participated in the evaluation said

they were satisﬁed or very satisﬁed with the recom-

mendations received. Together these two groups rep-

resent 57 students, 90.5% of those who responded

to the evaluation. Finally, the students had the op-

POI-based Recommender System for the Support of Academics in a Smart Campus

403

Figure 3: Usability and User satisfaction with the application.

Figure 4: Utility of the system.

portunity to leave suggestions to AONDE improve-

ments. Among the collected ideas are to present the

geographical location of the POIs, as if in the form

of a map or using a GPS system in real time. In addi-

tion, several students showed interest in learning more

about POIs, so that each one had more details. An-

other aspect raised was the possibility of knowing the

value of similarity for each recommended item, in the

same way creating gamiﬁcation strategies or integrat-

ing the application to social networks.

4 CONCLUSION

Recommendation systems (RS) are artifacts used to

manipulate large data sets, in order to reduce their

complexity by identifying and suggesting to users

only items of most interest or need. These software

tools seek to reduce the information burden while they

want to extract elements relevant to the context of use.

In the case of an environment integrated by an expres-

sive set of information, a Smart Campus proves to be a

suitable space for the application of an RS. Therefore,

an RS was proposed in the domain of the University,

whose objective is to support the students in their aca-

demic journey, relieving their cognitive and informa-

tional burden, with regard to the recommendation of

Points of Interest (POI) within the university campus.

After deepening and understanding about RS, the

proposal of this work was conceived, mainly aligned

with the vision of its target audience, since this project

is concerned with the use experience that it will offer.

Personas were drawn, representing the user’s proﬁle

in the Smart Campus domain, resulting in a content-

based POI Recommender System, linked to a web ap-

plication called AONDE.

The proposed algorithm represents the users and

the items to be recommended through tags, that are

keywords that describe such an entity. In this way,

both users and items are structured in matrices and

weighted by the TF metrics.

An interesting point of the RS developed in this

work is its concern in recommending items that are

necessary for a student in his/her ﬁrst steps at the uni-

versity, the freshman. A bias is applied to the user’s

proﬁle, considering this stereotype, so that items fun-

damentally necessary for this type of student receive

a greater weight.

Finally, the similarity between items and user is

calculated, using the cosine similarity, and generate

the outputs for recommendation. During the devel-

CSEDU 2020 - 12th International Conference on Computer Supported Education

404

Figure 5: User satisfaction with the recommendations received.

opment phase, several tests were carried out in or-

der to align users’ expectations. In addition, a pilot

test was conducted before the system was released

to the public. University students were able to inter-

act with AONDE, personalize their proﬁle and receive

POI recommendations within the campus. In the pe-

riod of the experiment, 110 users were created, which

enabled the present project to collect sufﬁcient data

for analysis.

Having the accuracy of RS and user satisfac-

tion as metrics for evaluation, the RS was evaluated.

Through a feedback system present in each recom-

mended item and a satisfaction questionnaire, evi-

dence was collected that guarantees the beneﬁts of the

system. It can be concluded that the system corre-

sponds to the desires of its target audience and actu-

ally contributes to the academic journey of a student

on campus, since 90.5% of the students say they felt

satisﬁed or very satisﬁed with the recommendations

received. Still on this evaluative perspective, the sys-

tem presented an accuracy in the recommendations

with an accuracy of 61% and a recall of 66%.

As future work the Recommender System can be

enhanced, since other recommender approaches could

be included in the system. For example, as more stu-

dents use the system on a regular basis, the collabora-

tive approach can be combined with content ﬁltering.

This work offers the opportunity to connect several

other elements that may contribute to its dissemina-

tion and use in the university community, such as a

gamiﬁcation system, whose goal would be to create

mechanisms of engagement and contribution to the

platform, disseminating knowledge and sharing expe-

riences in a playful and natural way.

ACKNOWLEDGEMENT

The authors would like to thank the funding pro-

vided by FAPESC (public call FAPESC/CNPq No.

06/2016 support the infrastructure of CTI for young

researchers, project T.O. No.: 2017TR1755 - Am-

bientes Inteligentes Educacionais com Integrac¸

ao de

ecnicas de Learning Analytics e de Gamiﬁcac¸

ao)

and UNIVERSAL MCTI/CNPq N

01/2016 -

Recomendac¸

ao adaptativa para cidades inteligentes.

This study was ﬁnanced in part by the Coordenac¸

de Aperfeic¸oamento de Pessoal de N

ıvel Superior -

Brasil (CAPES) Finance Code 001.

REFERENCES

Cantador, I., Bellog

ın, A., and Vallet, D. (2010). Content-

based recommendation in social tagging systems. In

Proceedings of the fourth ACM conference on Recom-

mender systems, pages 237–240. ACM.

Jannach, D., Zanker, M., Felfernig, A., and Friedrich, G.

(2010). Recommender Systems an Introduction. Cam-

bridge University Press, Leiden.

Lee, V., Schneider, H., and Schell, R. (2004). Mobile Ap-

plications: Architecture, Design, and Development.

Prentice Hall PTR, Upper Saddle River.

Machado, G. M., Maran, V., Dornelles, L. P., Gasparini,

I., Thom, L. H., and de Oliveira, J. P. M. (2018).

A systematic mapping on adaptive recommender ap-

proaches for ubiquitous environments. Computing,

100(2):183–209.

Pereira, R. and da Silva, S. R. P. (2008). Folksonomias:

Uma an

alise cr

ıtica focada na interac¸

ao e na natureza

da t

ecnica. In Proceedings of the VIII Brazilian Sym-

posium on Human Factors in Computing Systems,

IHC ’08, page 126–135, BRA. Sociedade Brasileira

de Computac¸

ao.

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

to recommender systems handbook. In Recommender

systems handbook, pages 1–35. Springer.

Ricci, F., Rokach, L., and Shapira, B. (2015). Recom-

mender systems: introduction and challenges. In Rec-

ommender systems handbook, pages 1–34. Springer.

POI-based Recommender System for the Support of Academics in a Smart Campus

405