RECAID: A Sponsorship Recommendation Approach

William Johnny Bernardes de Oliveira and Wladmir Cardoso Brand

Department of Computer Science, Pontiﬁcal Catholic University of Minas Gerais (PUC Minas), Belo Hozizonte, Brazil

Keywords:

Recommendation Systems, Recommender, Machine Learning, Supervised Learning, Sponsorship, Social

Project.

Abstract:

Non-government organizations play an important role in society, providing access to basic services in culture,

education, health, and security for needy people. Some of these organizations raise funds for their social

projects through sponsorship programs for people in poverty, deprivation, exclusion and vulnerability. The

intensive use of technology for sponsors and beneﬁciaries matching is paramount to create more lasting bonds,

maximizing the likelihood of stronger relationships, consequently raising more resources for projects. In this

article we propose and evaluate a learning approach to recommend beneﬁciaries to sponsors. Particularly, we

exploit different recommendation strategies, such as collaborative ﬁltering with matrix factorization, content-

based with bag of words and word embeddings and knowledge-based with association rules. Experimental

results show that content-based strategies based on word embeddings are more effective, reaching up to 72%

of performance in MAP and nDCG. Additionally, it can effectively recommend beneﬁciaries to sponsors even

if there is less feedback information on beneﬁciaries and sponsors to train recommendation models.

1 INTRODUCTION

Socioeconomic inequality is a worldwide problem.

While a large part of economic resources is exclu-

sively available to a small group of people, a large

group of people have no access to basic resources for

education and health. Particularly in Brazil, this prob-

lem is even greater. According to IPEA (Institute for

Applied Economic Research), approximately 26 mil-

lion people lived in poverty in 2014 in Brazil, and an-

other 8 million people lived in extreme poverty.

Many non-government organizations (NGOs)

support the social development of communities, de-

livering an improvement in the quality of life of

poor people. The Resolution 288 of the Economic

and Social Council of the United Nations deﬁnes

“non-government organizations” as organizations es-

tablished by civil society without government agree-

ments (Ferreira, 2005). NGOs often raise funding

from civil society and government to support their so-

cial development actions, but many of them have no

formal strategy for this (da Silva et al., 2016). Nev-

ertheless, one of the strategies adopted by NGOs is to

raise funds directly from individual sponsors through

sponsorship programs, where donors sponsor bene-

ﬁciaries by providing ﬁnancial resources, fostering

https://orcid.org/0000-0002-1523-1616

projects in the community where beneﬁciaries live.

Attracting and retaining sponsors is a challenging

problem and is closely related to the NGOs credi-

bility and theirs ability to effectively allocate fund-

ing resources. Previous work reported in literature

show that there is a positive correlation between trans-

parency in the application of funds and ﬁnancing

maintenance by sponsors (Portulhak et al., 2016).

Furthermore, emotional and practical aspects related

to the bonds between sponsor and beneﬁciary signif-

icantly impact the maintenance of donations. For in-

stance, emotional bonds present in the formal com-

munication between sponsor and beneﬁciary maxi-

mize the chance of lasting sponsorship. In this sce-

nario, attracting and retaining sponsors for NGOs

projects is paramount to the reduction of global so-

cioeconomic inequality. Thus, the understanding of

the emotional and practical aspects that impact the re-

lationship between sponsors and beneﬁciaries is cru-

cial to promote the creation of long-lasting bonds,

maximizing the likelihood of long-term funding.

In this article we propose a learning approach for

beneﬁciary-sponsor recommendation to retain spon-

sors through the building and maintenance of lasting

bonds between them. Particularly, different strategies

are used in recommendation, such as collaborative ﬁl-

tering with matrix factorization, content-based with

bag of words and word embeddings, and knowledge-

618

Bernardes de Oliveira, W. and Brandão, W.

RECAID: A Sponsorship Recommendation Approach.

DOI: 10.5220/0010400906180625

In Proceedings of the 23rd International Conference on Enter prise Information Systems (ICEIS 2021) - Volume 1, pages 618-625

ISBN: 978-989-758-509-8

based with association rules. Experimental results us-

ing a new sponsorship recommendation dataset built

from funding data provided by ChildFund Brasil

, an

international NGO, show that content-based strategies

with word embeddings outperforms the other strate-

gies, reaching up to 72% in MAP and nDCG. Ad-

ditionally, the content-based strategies performs well

even if there is less feedback information on bene-

ﬁciaries and sponsors to train the recommendation

model. The main contributions of this article are:

• SRD, a new sponsorship recommendation dataset

for beneﬁciary-sponsor recommendation built

from funding data of ChildFund Brasil, an inter-

national NGO.

• RECAID, a learning approach to recommend ben-

eﬁciaries to sponsors that implements different

recommendation strategies, particularly a content-

based strategy driven by word embeddings.

• A throughout evaluation of our proposed ap-

proach, contrasting the performance of different

recommendation strategies.

This article indirectly contributes to improve the NGO

fundraising process, presenting concepts and tech-

nologies to enhance NGO processes and methods.

Moreover, our recommendation approach is designed

to be effective in a real-world scenario where most

beneﬁciary-sponsor relationships are one to one. Un-

like other recommendation approaches that recom-

mend knowing a large range of choices made by

users, RECAID need to be effective in a more restrict

universe of knowing choices.

The remaining of this article is divided as follows:

Section 2 presents theoretical background on social

technologies and recommendation systems. Section

3 presents related work. Section 4, presents our pro-

posed recommendation approach. Section 5 presents

experimental setup and results. Finally, Section 6

presents conclusions and directions for future work.

2 BACKGROUND

This section presents the main concepts on social

technologies and recommendation systems.

2.1 Social Technologies

The expression “social technology” originates in In-

dia from a concept of appropriate technology, used

for the ﬁrst time in the late 14th century by Mahatma

http://www.childfundbrasil.org.br

Gandhi (Dagnino et al., 2004). In Brazil, this ex-

pression became popular in the ﬁrst decade of the

21st century, mostly due to social actors concerned

with the problem of social exclusion. Formally, so-

cial technology is a way to use, create, implement

and manage technologies to address social and envi-

ronmental problems, trying to promote social inclu-

sion and sustainable development in a guided way,

comprising products, techniques, and methodologies

developed for community interaction with the aim

of promote social transformation (Dagnino, 2011).

Thus, social technologies require action and reﬂec-

tion of individuals to stimulate a more fair, inclusive

and sustainable society, emphasizing citizens, neigh-

borhood associations, solidarity economy enterprise,

social mobilization and organizations (Costa, 2013).

Typically, social technologies run by social orga-

nizations action within social projects. These organi-

zations usually have no links with the public or private

sector, do not receive government funding and their

projects are sponsored by donations from ordinary

contributors (Falc

ao, 2004; de Albuquerque, 2006).

The two key agents in social projects are beneﬁcia-

ries and sponsors. Beneﬁciaries are people who are

directly beneﬁt from the project’s products, services

and results. Usually, they are needy people and their

needs or rights are not met by public agencies (Cul-

ligan et al., 2013). Sponsors are donors in social

projects that provide cash, goods or services to the

NGOs or directly to a non-proﬁt social project. These

sponsors contribute by providing more opportunities

and resources for beneﬁciaries (Belem and Donadone,

2013).

2.2 Recommendation Systems

Recommendation systems, or recommenders, are in-

formation retrieval systems that help users to make

choices on items in a wide universe of options. Rec-

ommendations are generated through explicitly or im-

plicitly preferences expressed by users, and by using

items and users properties, including demographic

and lifestyle data, visualization statistics, behavioral

information and other contextual properties depend-

ing on what is recommend (Linden et al., 2003; Massa

and Avesani, 2007; Cremonesi et al., 2010).

Until recently recommendation was a personal ex-

perience based on people habits and behaviors. From

there, it has been evolving to a personal experience

based on collective knowledge. Today, there are a

huge amount of choices of items in the Web, such as

movies, music, food, drinks and other products, mak-

ing decisions hard for people emerged in their daily

routines. Recommenders offer a way to streamline

RECAID: A Sponsorship Recommendation Approach

619

decisions, ﬁltering the options based on past choices,

common interests and preferences.

Popularity-based is one of the most traditional rec-

ommendation strategy widely used in practise, where

the only available knowledge used to recommend is

the history and popularity of the items. This strategy

is based on collective wisdom and typically recom-

mend the most popular items to users, with no cus-

tomization or personalized experience (Steck, 2011).

It does not distinguish groups, providing no targeted

recommendations in such a way that it is often inef-

fective, failing to embody the user’s desire.

The collaborative ﬁltering strategy is more sophis-

ticated, providing predictions aimed at the users’ in-

terest on items, consolidating ﬁltering based on their

preferences history (Linden et al., 2003). These

strategies assume the likelihood of users having com-

mon interests about items, being able to group users

based on the grade they have deﬁned for an item in

the past, and infer the user’s future choice by un-

derstanding the preferences that similar users in the

same group made for certain items (Sarwar et al.,

2001). Usually, collaborative ﬁltering recommenda-

tion is based on the similarity of users (user-based)

or on the similarity of items (item-based). In addi-

tion, latent models (Sarwar et al., 2000; Aggarwal

and Parthasarathy, 2001) reduces the dimensionality

inherent in calculating similarity in large matrices of

users and items, providing an effective matrix factor-

ization procedure. One of the most popular imple-

mentations of latent models is the singular value de-

composition (SVD).

The content-based recommendation strategy uses

the textual content of users and items to provide pre-

dictions. It can recommend new items for users even

when there is no feedback data, minimizing the cold-

start problem. However, due to the use of text de-

scriptors, it usually provides obvious recommenda-

tions (Balabanovi

c and Shoham, 1997; Aggarwal

et al., 2016). A hybrid strategy can achieve out-

performing recommendation, improving other non-

hybrid strategies (Lekakos and Caravelas, 2008). It

aims to minimize problems related to cold-start and

sparsity by combining different strategies, depending

on what kind of item need to be recommended. The

hybrid strategy is widely used nowadays due to its ef-

fectiveness, resilience and modularity.

Knowledge-based recommendation strategies use

explicit contextual knowledge on items, user prefer-

ences and recommendation criteria to provide predic-

tions (Burke, 2000). Usually these strategies are suit-

able for complex domains, where items are not con-

sumed very often by users. Several classiﬁer and re-

gression algorithms can be used to provide the pre-

dictions, such as association rules (Osadchiy et al.,

2019), support vector machines (Min and Han, 2005)

and neural networks (Gupta and Sharma, 2021). Par-

ticularly, a recommendation strategy based on neural

networks is inspired by the architecture of the human

brain, allowing for an important ability to train differ-

ent classiﬁers for a higher quality recommendation.

Similar to hybrid recommenders, neural network rec-

ommenders can combine several architectures and be

trained using several data sources to provide an out-

performing recommendation. Thus, one can train sev-

eral neural networks per user, still customizing the

user preferences, tastes, and behavior (Ricci et al.,

2015).

Recently, novel deep learning architectures have

been proposed to generate text embeddings that can

be effectively used in several natural language pro-

cessing problems (Mikolov et al., 2013b; Vaswani

et al., 2017; Devlin et al., 2019). For instance,

Word2Vec are the ﬁrst efﬁcient models to learn dis-

tributed representations of words from large amount

of unstructured text with billions of words (Mikolov

et al., 2013a; Mikolov et al., 2013b). Training such

models does not involve dense matrix multiplications

and can be quickly done with a single machine. Pre-

vious work reported in literature show that word em-

beddings can be effectively used in recommenda-

tion (Kannan et al., 2018).

Regardless of strategy, recommendation systems

must provide personalized recommendations by us-

ing as much information as possible on demographic,

lifestyle, groups of friends, areas of interest and any

other feature that can deﬁne user personality. Pro-

viding effective recommendations is paramount to re-

tain users, particularly when they are able to convey a

sense of individuality, that is when the user feels that

their recommendations are exclusive. By analogy, an

effective recommendation system can be associated

with an experienced salesperson, that ﬁrst seeks to un-

derstand the customer’s needs to then recommend a

product that meet the needs. In this sense it is crucial

to know historical data and characteristics of users to

provide effective recommendations (Zanker and Nin-

aus, 2010).

3 RELATED WORK

Recommending people (items) to people (users) can

be applied in different contexts, such as friendships,

dating, educational and professional partnerships. In

this context, users lifestyle information is usually an

important feature for recommendation and, together

with others features, it can provide an approximation

ICEIS 2021 - 23rd International Conference on Enterprise Information Systems

620

of the user’s personality, allowing the predictions of

user future choices to be ﬁltered in a better way.

For friendship recommendation, (Gurini et al.,

2018) proposed a new recommendation approach

based on semantic attitudes by using traditional so-

cial networks, such as Twitter. In particular, favorable

and unfavorable statements are considered in speciﬁc

matters, thus improving the recommendation results.

They extract text from social networks, converting

them into unigrams and eliminating irrelevant ones. A

three-dimensional matrix factorization model is gen-

erated and a temporal dynamics is modeled, improv-

ing the accuracy and diversity of the recommender.

The experiments carried out by authors with data ob-

tained in the monitoring of the trafﬁc produced by

the users allowed the authors to make comparative

analyzes with other approaches reported in the liter-

ature, such as random recommendation, popularity-

based, content-based, and collaborative ﬁltering. The

authors divided the datasets into 70% for training and

30% for testing, for the evaluation of results and ac-

curacy, diversity and novelty metrics were measured.

Experimental results showed that with the use of im-

plicit feeling, volume and objectivity improves the

recommendations. In addition, the proposed approach

behaved better than the others in experimental evalu-

ation. Similarly to these work, in our investigation

we analyse features that impact the recommendation

of people to people. Moreover, we also use a consoli-

dated approach in the literature to try out possibilities.

But differently from these work, we use a sponsor-

ship recommendation dataset to infer the likelihood of

long-term bond between beneﬁciaries and sponsors.

In a different vein, (Rahim et al., 2019) propose

an approach based on centrality measures between

friends of friends and similarity measures of groups

and events to recommend new friends for users. The

features used by this approach is extracted from two

Facebook datasets, the ﬁrst consisting of seven group

of features: friends, events, groups, neighbors, friend

events, friend event groups and friend event groups

neighbors. The second consisting on friendship rela-

tions and used to assess whether both users were able

to become true friends. For recommendation, the au-

thors propose different techniques for measuring sim-

ilarities. The ﬁrst technique is based on friends of

friends relationships, with the ﬁve best friends been

recommended using each one of the centrality mea-

sures separately. The second technique use three

measures of similarity based on each user’s common

group or event: data coefﬁcient, Jaccard similarity,

and cosine similarity. Both techniques were evaluated

and the authors showed that the recommendations

made by them performs almost equally, with an ac-

curacy of 56% for the second technique and 52% for

the ﬁrst technique. In addition, these work presents an

interesting comparison between techniques and met-

rics, creating possibilities for the authors to verify the

conﬁdence level of the recommendation.

The aforementioned related approaches that rec-

ommend people focus mainly on the interest of new

friendships or dating. In the same vein (Edith and

Yu, 2018) propose a recommender focused on the

building of new friendships. The authors use the k-

nearest neighbors algorithm to estimate users’ prefer-

ences and lifestyle features. They also collected data

directly from each user’s smartphones, which allowed

measuring the similarity between them having a de-

ﬁned context. Experimental results showed that the

friends’ choices directly reﬂect the preferences ex-

pressed by the users, with the main advantage in rec-

ommending friends, the sharing of similar interests

on both sides. This work is important to demonstrate

how similarities measures of users impact the recom-

mendation of friends, and how estimating similar in-

terests, daily routines, styles, and opinions can help to

build users clusters that enhance the recommendation.

Although there are several works reported in the

literature on recommending people, none of them ad-

dress the problem of recommending beneﬁciaries to

donors, a sub-problem that imposes even greater chal-

lenges given the properties of the users and items of

the recommendation system, and the lack of informa-

tion on social sponsorship.

4 THE RECAID APPROACH

Recommendation systems can be implemented in

different contexts using different recommendation

strategies. In this section we present RECAID, our

approach to recommend beneﬁciaries to sponsors in

social projects. Figure 1 presents the RECAID archi-

tecture.

First, the Processor component creates the Spon-

sorship dataset by processing the NGO Funding

dataset containing sponsor and beneﬁciary relation-

ships. Part of this processing consists of anonymizing

the data of beneﬁciaries and sponsors. Another part

of processing consists of generating sponsor ratings

for beneﬁciaries based on the level of interactions be-

tween them. Each case record in the NGO Funding

dataset represents an interaction between a sponsor

and a beneﬁciary. In particular, the dataset records

eleven types of cases. In one year, at least three cases

are record for each beneﬁciary: progress report, reg-

istration update, and initiated letter. Some engaged

sponsors have a larger number of cases, usually hav-

RECAID: A Sponsorship Recommendation Approach

621

Figure 1: The architecture of the RECAID approach.

ing a longer time bond with their beneﬁciaries.

Second, the Extractor component extracts associa-

tion rules with high conﬁdence level (above a conﬁg-

urable threshold) from the Sponsorship dataset. For

this, we use sponsor’s features, such as city, gender,

age and payment, and beneﬁciary’s features as gen-

der, age and illiteracy level. Third, the Recommender

component learns from the Sponsorship dataset and

association rules a recommendation model using dif-

ferent strategies. In particular, collaborative ﬁltering,

content-based, knowledge-based and hybrid strate-

gies are used for this. Finally, the Predictor com-

ponent recommend a ranked set of beneﬁciaries to a

sponsor using the recommendation model.

5 EXPERIMENTS

In this section, we present the experiments we car-

ried out to evaluate our recommendation approach,

including datasets, evaluation metrics and experimen-

tal results. Particularly, in our experiments we answer

the following research questions: how effective is RE-

CAID to recommend beneﬁciaries to sponsors in so-

cial projects?

5.1 Dataset

The sponsorship recommendation dataset (SRD)

proposed in this article was built using funding data

provided by ChildFund Brasil, an international NGO.

This new dataset contains data of 53 years of so-

cial projects development, ﬁlling a gap in the lack of

http://doi.org/10.5281/zenodo.4540776

datasets for beneﬁciary-sponsor recommendation sys-

tems. Within the funding data we extract seventeen

sources of features that bring information on beneﬁ-

ciaries and sponsors, such as their demographic and

personal characteristics.

Along with the SRD dataset we provide a short-

code showing how to use it for beneﬁciary-sponsor

recommendation, in addition to sponsorship meta-

data in English and Portuguese. In particular, the

SRD dataset contains 11,392 records of beneﬁciaries,

8,624 of them already being sponsored by any spon-

sor. In addition, there are 16,923 records of spon-

sors, 8,299 of them donating infrequently. Moreover,

68% of beneﬁciary-sponsor links are unique, i.e., one

donor sponsor just one beneﬁciary.

In our experiments, we use four sources of

beneﬁciary-sponsor features: beneﬁciaries, bonds,

cases, and sponsors. To infer sponsor ratings in ben-

eﬁciaries, we perform the sum of all cases, i.e., the

interactions recorded in the sponsorship relationship,

and we normalize each beneﬁciary-sponsor case by

the sum of cases.

5.2 Setup

We evaluate four different strategies for recom-

mendation: collaborative ﬁltering, content-based,

knowledge-based and a hybrid one. For collabora-

tive ﬁltering we use the cosine similarity as the dis-

tance metric used to calculate sponsors and beneﬁcia-

ries similarity and matrix factorization with singular

value decomposition. In our experiments, we refer to

this strategy as SVD.

For the content-based strategy, all sponsors and

beneﬁciaries are represented in a vector space us-

ing bag of words with TF-IDF schema (BoW) and

ICEIS 2021 - 23rd International Conference on Enterprise Information Systems

622

Word2Vec representations. For both BoW and

Word2Vec we estimate the similarity between the vec-

tors using the cosine similarity. We set the maximum

size of each vector of 5,000 words. These words are

extracted from the beneﬁciary and sponsor descrip-

tions. In our experiments, we refer to the content-

based strategy with TF-IDF as CB and to the content-

based strategy with Word2Vec as W2V. We also com-

bine TF-IDF and Word2Vec vectors in a single vec-

tor. In our experiments, we refer to the content-

based strategy with combined TF-IDF and Word2Vec

as W2VT.

For the hybrid strategy, we multiply the scores

from collaborative ﬁltering and content-based strate-

gies previously presented to provide a hybrid rank-

ing. In our experiments, we refer to the hybrid strat-

egy as HB. Additionally, we combine previously pre-

sented strategies with a knowledge-based recommen-

dation strategy based on association rules, referring

them in our experiments as CB-AR, HB-AR, SVD-

AR, W2V-AR and W2VT-AR. We extract all rules

with conﬁdence levels between 20% and 75%, ﬁlter-

ing from the recommended beneﬁciaries only those

records that follow the association rules.

There are different metrics to evaluate the accu-

racy of a recommendation system (Beel et al., 2013).

In this article, we provide ranking prediction metrics

to evaluate the order in which the items should be

recommended. For this, we use information retrieval

metrics, such as, Mean Average Precision (MAP) and

Normalized Discounted Cumulative Gain (nDCG).

The MAP calculates the average precision of the

scores of a set of queries and the nDCG deﬁnes the

relevance of the classiﬁcation of the items recom-

mended in the result set to evaluate the utility or gain

based on its position (Geng et al., 2007). In addition,

we use the holdout for cross-validation, considering a

random sample of 20%, 30%, 40%, 50%, 60% train-

ing data,

5.3 Experimental Results

Table 1 presents the MAP metric for each recommen-

dation strategy.

From Table 1 we observe that W2VT outperforms

the other approaches. However the performance of

HB, HB-AR, and SVD are remarkable. The perfor-

mance of HB and HB-AR is largely derived from the

performance of the SVD. Additionally, we observe

that the best performance occurs when the sample

rate is 40% for training and 60% for testing. Table 2

presents the nDCG metric for each strategy.

From Table 2 we observe again that W2VT out-

performs the other approaches. The CB strategy is the

Table 1: RECAID Performance in MAP.

Approach 20% 30% 40% 50% 60%

CB 0.349 0.313 0.351 0.128 0.048

CB-AR 0.256 0.224 0.173 0.001 0.002

HB 0.555 0.428 0.500 0.500 0.666

HB-AR 0.555 0.428 0.500 0.500 0.666

SVD 0.556 0.428 0.500 0.501 0.666

SVD-AR 0.222 0.142 0.333 0.250 0.000

W2V 0.333 0.364 0.329 0.383 0.322

W2V-AR 0.210 0.235 0.110 0.197 0.181

W2VT 0.580 0.480 0.598 0.568 0.721

W2VT-AR 0.398 0.428 0.351 0.343 0.390

Table 2: RECAID Performance in NDCG.

Approach 20% 30% 40% 50% 60%

CB 0.666 0.571 0.500 0.250 0.000

CB-AR 0.333 0.285 0.166 0.000 0.000

HB 0.555 0.428 0.500 0.500 0.666

HB-AR 0.555 0.428 0.500 0.500 0.333

SVD 0.556 0.428 0.500 0.500 0.666

SVD-AR 0.222 0.142 0.333 0.250 0.000

W2V 0.556 0.428 0.500 0.500 0.666

W2V-AR 0.322 0.242 0.433 0.250 0.410

W2VT 0.708 0.651 0.612 0.668 0.721

W2VT-AR 0.708 0.480 0.612 0.668 0.601

most volatile, but with minor differences when com-

pared. The association rules extracted from the base,

also do not help in gaining the correctness of the strat-

egy as it was expected to have, we can see, at most in

the hybrid strategy, the hits are close, but they do not

exceed.

The experimental results show that W2VT out-

performs the other strategies in all evaluation met-

rics. Recalling our research question, these obser-

vations attest the effectiveness of RECAID to recom-

mend beneﬁciaries to sponsors in social projects, es-

pecially when we use features that describe a bene-

ﬁciary, that is, when we opt for content-based strate-

gies.

The content-based strategy that uses the descrip-

tion or attributes of the beneﬁciaries and thus creates

similarities was more efﬁcient. To use this strategy,

we used the description ﬁeld of the beneﬁciary, us-

ing the TF-IDF scheme to calculate the weight of

the words in the text, which is a measure to indi-

cate the importance of each word in the description

of the beneﬁciary, and with that, we can verify the

links and determine the most similar ones helping to

solve the problem we have in the other recommen-

dations to which they present only a 1 to 1 ratio in

most cases. The best results come from the TF-IDF

Word2Vec, Pattern Factorization and Hybrid strategy,

combining two strategies already presented, based on

content and pattern factoring (SVD), thereby mini-

mizing problems related to sparsity and cold-start.

RECAID: A Sponsorship Recommendation Approach

623

By assessing the effectiveness of these strategies,

we noticed that the sponsor and beneﬁciary relation-

ship structure greatly inﬂuences the models, keeping

them away from improved results, since the history

or similarity that are important factors for a good rec-

ommendation establish little performance. However,

when evaluating the strategies that had a better result,

we realized that the description of the beneﬁciary can

establish a path to be followed in the recommenda-

tion, ﬁnding similarities between them. For the tests,

we do not consider the temporal dimension and we

consider that all beneﬁciaries would be available to

receive a sponsor in the sponsorship program. It is

also not possible to discard other unused base fea-

tures, which can improve recommendation strategies.

6 CONCLUSION

In the present article we propose RECAID, a recom-

mendation approach that recommends beneﬁciaries to

sponsors in social projects. RECAID aims to improve

fundraising and loyalty from sponsors through the

construction of lasting bonds. In particular, different

recommendation strategies reported in scientiﬁc liter-

ature on recommender systems for people recommen-

dation were identiﬁed, different techniques for recom-

mending beneﬁciaries to sponsors in social projects

were evaluated and the main characteristics of bene-

ﬁciaries and sponsors that impact on the effectiveness

of the proposed recommendation approach were iden-

tiﬁed.

Experimental results show that the Wor2vec with

TF-IDF that exploits descriptive textual content of

beneﬁciaries and sponsors performed better for rec-

ommendation. Particularly, with a training rate of

60%, the MAP and nDCG metrics reached 72% ac-

curacy. In addition, other as matrix factorization and

hybrid approaches strategy achieved the accuracy of

66% in nDCG. Based on the observed results we in-

tend to explore new techniques for displaying the tex-

tual content and new features that describe emotional

aspects of beneﬁciaries and sponsors.

In future work, we also intend to explore a larger

set of characteristics of beneﬁciaries and sponsors to

understand what practical and emotional aspects im-

pact the maintenance of lasting bonds. Additionally,

we intend to evaluate other strategies recently pro-

posed in the literature, such as those based on neu-

ral networks. Moreover, we intend to use a larger

set of features, since the current database is unbal-

anced, with a small number of lasting links between

beneﬁciaries and sponsors. We also intend to ex-

plore new approaches based on sentence embeddings

to capture neglected semantic aspects to improve the

performance of content-based strategies.

ACKNOWLEDGEMENTS

The present work was carried out with the support

of the Coordenac¸

ao de Aperfeic¸oamento de Pessoal

de N

ıvel Superior - Brazil (CAPES) - Financing

Code 001. The authors thank the partial support of

the CNPq (Brazilian National Council for Scientiﬁc

and Technological Development), FAPEMIG (Foun-

dation for Research and Scientiﬁc and Technological

Development of Minas Gerais), and PUC Minas.

REFERENCES

Aggarwal, C. C. et al. (2016). Recommender systems.

Springer.

Aggarwal, C. C. and Parthasarathy, S. (2001). Mining mas-

sively incomplete data sets by conceptual reconstruc-

tion. In Proceedings of the 7th ACM SIGKDD Interna-

tional Conference on Knowledge Discovery and Data

Mining, KDD’01, pages 227–232.

Balabanovi

c, M. and Shoham, Y. (1997). Fab: content-

based, collaborative recommendation. Communica-

tions of the ACM, 40(3):66–72.

Beel, J., Langer, S., Genzmehr, M., Gipp, B., Breitinger,

C., and N

urnberger, A. (2013). Research paper rec-

ommender system evaluation: a quantitative literature

survey. In Proceedings of the International Workshop

on Reproducibility and Replication in Recommender

Systems Evaluation, RepSys’13, pages 15–22.

Belem, M. P. and Donadone, J. C. (2013). The Rouanet Law

and the construction of a cultural sponsor market. New

Sociological Directions, 1(1).

Burke, R. (2000). Knowledge-based recommender systems.

Encyclopedia of Library and Information Systems.

Costa, A. B. (2013). Social technology & public policies.

Instituto P

olis.

Cremonesi, P., Koren, Y., and Turrin, R. (2010). Perfor-

mance of recommender algorithms on top-n recom-

mendation tasks. In Proceedings of the 4th ACM Con-

ference on Recommender Systems, RecSys’10, pages

39–46.

Culligan, M., Marks, S., Nelson, T., Radstone, L., and

Verzuh, E. (2013). A guide to the PMD Pro.

da Silva, E. P. C., de Vasconcelos, S. S., and Nor-

manha Filho, M. A. (2016). Third sector organiza-

tions: Challenges in raising funds for their manage-

ment. Qualis Sumar

e - Electronic Academic Maga-

zine, 6(2).

Dagnino, R. (2011). Social technology: conceptual basis.

Science & Social Technology, 1(1):1–12.

Dagnino, R., Brand

ao, R., and Novaes, H. (2004). So-

cial technology: A strategy for development. Rio de

Janeiro: Banco do Brasil Foundation, page 209.

ICEIS 2021 - 23rd International Conference on Enterprise Information Systems

624

de Albuquerque, A. C. C. (2006). Third sector: history and

management of organizations. Summus Editorial.

Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K.

(2019). BERT: Pre-training of deep bidirectional

transformers for language understanding. In Proceed-

ings of the 2019 Conference of the North American

Chapter of the Association for Computational Lin-

guistics: Human Language Technologies, NAACL-

HLP’19, pages 4171–4186.

Edith, M. B. S. and Yu, W. (2018). Friend recommendation

system based on mobile data. In Proceedings of the

International Conference on Engineering Simulation

and Intelligent Control, ESAIC’18, pages 326–329.

Falc

ao, J. (2004). Democracy, law and the third sector,

volume 1. Fundac¸

ao Get

ulio Vargas.

Ferreira, V. C. P. (2005). NGOs in Brazil: a study on

their characteristics and factors that have driven their

growth. PhD thesis, Fundac¸

ao Get

ulio Vargas.

Geng, X., Liu, T.-Y., Qin, T., and Li, H. (2007). Fea-

ture selection for ranking. In Proceedings of the 30th

Annual International ACM SIGIR Conference on Re-

search and Development in Information Retrieval, SI-

GIR’07, pages 407–414.

Gupta, A. and Sharma, A. (2021). Implementation of rec-

ommender system using neural networks and deep

learning. In Sharma, M. K., Dhaka, V. S., Perumal,

T., Dey, N., and Tavares, J. M. R. S., editors, Inno-

vations in Computational Intelligence and Computer

Vision, pages 256–263.

Gurini, D. F., Gasparetti, F., Micarelli, A., and Sansonetti,

G. (2018). Temporal people-to-people recommenda-

tion on social networks with sentiment-based matrix

factorization. Future Generation Computer Systems,

78:430–439.

Kannan, M. S., Mahalakshmi, G. S., Smitha, E. S., and

Sendhilkumar, S. (2018). A word embedding model

for topic recommendation. In Proceedings of the

2nd International Conference on Inventive Communi-

cation and Computational Technologies, ICICCT’18,

pages 826–830.

Lekakos, G. and Caravelas, P. (2008). A hybrid approach

for movie recommendation. Multimedia tools and ap-

plications, 36(1-2):55–70.

Linden, G., Smith, B., and York, J. (2003). Amazon.com

recommendations: Item-to-item collaborative ﬁlter-

ing. IEEE Internet computing, 1:76–80.

Massa, P. and Avesani, P. (2007). Trust-aware recommender

systems. In Proceedings of the ACM Conference on

Recommender Systems, RecSys’07, pages 17–24.

Mikolov, T., Chen, K., Corrado, G. S., and Dean, J. (2013a).

Efﬁcient estimation of word representations in vector

space. CoRR, abs/1301.3781.

Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S., and

Dean, J. (2013b). Distributed representations of words

and phrases and their compositionality. In Proceed-

ings of the 26th International Conference on Neu-

ral Information Processing Systems, NIPS’13, pages

3111–3119.

Min, S.-H. and Han, I. (2005). Recommender systems

using support vector machines. In Proceedings of

the 5th International Conference on Web Engineering,

ICWE’05, page 387–393.

Osadchiy, T., Poliakov, I., Olivier, P., Rowland, M., and

Foster, E. (2019). Recommender system based on

pairwise association rules. Expert Systems with Ap-

plications, 115:535–542.

Portulhak, H., Delay, A. J., and Pacheco, V. (2016). Ac-

countability by third sector entities and their impact

on obtaining resources: a look at the behavior of indi-

vidual donors. Think Accounting, 17(64).

Rahim, M., Ejaz, A., Rajput, Q., and Khoja, S. A. (2019).

A comparative study of similarity and centrality mea-

sures for friends recommendation. In Proceedings of

the International Conference on Information Science

and Communication Technology, ICISCT’19, pages

1–8.

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

mender systems: introduction and challenges. In Rec-

ommender systems handbook, pages 1–34. Springer.

Sarwar, B., Karypis, G., Konstan, J., and Riedl, J. (2000).

Application of dimensionality reduction in recom-

mender system-a case study. Technical report, Min-

nesota University.

Sarwar, B., Karypis, G., Konstan, J., and Riedl, J. (2001).

Item-based collaborative ﬁltering recommendation al-

gorithms. In Proceedings of the 10th International

Conference on World Wide Web, WWW ’01, page

285–295.

Steck, H. (2011). Item popularity and recommendation ac-

curacy. In Proceedings of the 5th ACM Conference on

Recommender Systems, RecSys’11, pages 125–132.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones,

L., Gomez, A. N., Kaiser, L., and Polosukhin, I.

(2017). Attention is all you need. In Proceedings of

the 31st Conference on Neural Information Process-

ing Systems, NIPS’17.

Zanker, M. and Ninaus, D. (2010). Knowledgeable explana-

tions for recommender systems. In Proceedings of the

IEEE/WIC/ACM International Conference on Web In-

telligence and Intelligent Agent Technology, volume 1

of WI-IAT’10, pages 657–660.

RECAID: A Sponsorship Recommendation Approach

625