Exploring Strategies to Mitigate Cold Start in Recommender Systems: A

Systematic Literature Mapping

Nathalia Locatelli Cezar

1 a

, Isabela Gasparini

1 b

, Daniel Lichtnow

2 c

Gabriel Machado Lunardi

2 d

and Jos

e Palazzo Moreira de Oliveira

3 e

Universidade do Estado de Santa Catarina (UDESC), R. Paulo Malschitzki 200, Joinville, Brazil

Universidade Federal de Santa Maria (UFSM), Av. Roraima 1000, Santa Maria, Brazil

Universidade Federal do Rio Grande do Sul (UFRGS), Porto Alegre, Brazil

Keywords:

Recommender Systems, Cold Start, New User Cold Start.

Abstract:

Recommender Systems are designed to provide personalized item recommendations to users based on their

preferences and behavioral patterns, aiming to suggest items that align with their interests and proﬁle. In

Recommender Systems, a common issue arises when the user’s proﬁle is not adequately characterized, par-

ticularly at the initial stages of using the system. This issue has persisted in Recommender Systems since its

inception, commonly known as Cold Start. The Cold Start issue, which impacts new users, is called User Cold

Start. Through a systematic literature mapping, this paper identiﬁes strategies to minimize User Cold Start

without reliance on external sources (such as social networks) or user demographic data for initializing the

proﬁle of new users. The systematic literature mapping results present strategies aimed at mitigating the User

Cold Start Problem, serving as a foundational resource for further enhancements or novel proposals beyond

those identiﬁed in the review. Thus, the goal of this work is to understand how to create an initial user proﬁle

before any prior interaction and without using external sources in the recommender system.

1 INTRODUCTION

Recommender systems provide suggestions for items

with a higher probability of interest to a user (Ricci

et al., 2015). An “item” refers to what the system

recommends to users (Jannach et al., 2010), which

can be movies, books, music, tourist spots, products,

scientiﬁc papers, etc.

Since the inception of Recommender Systems,

various algorithms/approaches have been devised to

produce recommendations. Traditionally, three ap-

proaches are mainly cited: (a) Content-Based, where

the user receives recommendations for items similar

to those they preferred in the past; (b) Collaborative

Filtering, where the system recommends items to the

user that users evaluated with a similar proﬁle; and

mend items, aiming to reduce disadvantages present

in a single approach (Burke, 2002). Other approaches

can be cited, like the demographic approach where so-

https://orcid.org/0000-0003-2727-2121

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

https://orcid.org/0000-0003-0103-0538

https://orcid.org/0000-0001-6655-184X

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

ciodemographic attributes such as age are used (Ricci

et al., 2015).

Despite the advancements in Recommender Sys-

tems, with numerous published works and the ap-

plication of these systems in various domains, there

is still much to explore, especially regarding some

known problems of these systems (Alyari and Ja-

fari Navimipour, 2018). One of the most well-known

issues is the Cold Start problem. Despite the approach

employed, Recommender Systems grapple with the

Cold Start problem (Ricci et al., 2015). The Cold

Start problem can be deﬁned as the inability to create

reliable recommendations due to a lack of data about

a new user or a new item (Monti et al., 2021). Thus,

there are two types of Cold Start: User Cold Start and

Item Cold Start. New User Cold Start is related to the

fact that a user who has interacted little with the sys-

tem and made a few item evaluations will not initially

receive quality recommendations since little is known

about this user. New Item Cold Start refers to the fact

that items evaluated by few users or not evaluated at

all will not be recommended, something that occurs

in some of the approaches used in Recommender Sys-

tems, specially in Collaborative Filtering.

The Cold Start poses serious challenges to the

value of Recommender Systems, as users who do not

Cezar, N., Gasparini, I., Lichtnow, D., Lunardi, G. and Moreira de Oliveira, J.

Exploring Strategies to Mitigate Cold Start in Recommender Systems: A Systematic Literature Mapping.

DOI: 10.5220/0012550700003690

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 26th International Conference on Enterprise Information Systems (ICEIS 2024) - Volume 1, pages 965-972

ISBN: 978-989-758-692-7; ISSN: 2184-4992

965

receive appropriate recommendations for their proﬁle

may, in some cases, cease using the system (Panda

and Ray, 2022). In (Panda and Ray, 2022), the authors

emphasize that Cold Start problems represent serious

challenges to the commercial value of Recommender

Systems, where New users who do not ﬁnd the rec-

ommendations useful would stop using the system,

affecting user engagement and sales.

In the work of (Abdullah et al., 2021), the authors

conclude that little research has been carried out in

academia about the Could Start problem. However,

they found that cold start user recommendation has

frequently been researched in the entertainment do-

main, typically using music and movie data.

Thus, this work aims to identify attempts to min-

imize the Cold Start problem described in academic

literature through a systematic literature mapping.

The focus is on the User Cold Start problem. This

choice was made since User Cold Start is the more

crucial characteristic for user engagement and reten-

tion of new users, and it is not a highlighted prob-

lem in a more signiﬁcant number of approaches used

in Recommender Systems (see Section 2). Also, the

number of these approaches used to create the users’s

proﬁles using some prior interaction and using exter-

nal sources for the recommendation system is exten-

sive. Consequently, this paper brings a comprehen-

sive analysis to understand how to create this proﬁle

without those approaches.

This paper is structured as follows. Section 2

presents more details about the Cold Start problem,

along with papers describing how Recommender Sys-

tems has obtained data to identify user preferences

and build their proﬁles. Section 3 presents the system-

atic mapping carried out. Considering the results of

systematic mapping, Section 4 indicates future works.

Section 5 presents the ﬁnal considerations.

2 USER PROFILE

CONSTRUCTION AND COLD

START

The user cold start problem is linked to the creation

of user proﬁles. Understanding the process of cre-

ating user proﬁles requires knowledge about how to

build and represent these proﬁles. The user proﬁle

representation depends on the chosen approach within

Recommender Systems. It is also important the way

to gather data for building robust user proﬁles. The

subsequent sections explore these aspects.

2.1 User Proﬁle Representation and

Recommender Systems Approaches

Cold Start presents itself differently for each rec-

ommendation approach employed. In the Content-

Based approach, where the user receives recommen-

dations for items similar to those they liked in the

past, the problem is particularly pronounced for new

users ((New User Cold Start) (Pazzani and Billsus,

2007). In Collaborative Filtering, in addition to User

Cold Start, there is also the presence of New Item

Cold Start, which is related to the fact that items not

evaluated by any user will not be recommended (Lika

et al., 2014)

The issue lies in the Content-Based approach No-

tably, the Content-Based approach is closely linked

to the ﬁeld of Information Retrieval, where items are

often characterized by textual descriptions (papers,

books, etc.) (Adomavicius and Tuzhilin, 2005).

In the Content-Based approach, the user’s pro-

ﬁle is represented by a set of keywords (Adomavi-

cius and Tuzhilin, 2005). In many Recommender

Systems, this set of keywords in the user proﬁle is

compared with items (papers, books, for example) us-

ing a similarity function, like cosine. Furthermore,

in Content-Based Recommender Systems, TF-IDF -

Term Frequency-Inverse Document Frequency is em-

ployed (TF-IDF is frequently used in Information Re-

trieval Systems). The TF-IDF measures the impor-

tance of a term/word to a document in a collection.

The TF-IDF assigns a weight to term i in document

d. Here, t f

i,d

represents the number of occurrences of

term i in document d, and id f

, known as the Inverse

Document Frequency of term i, emphasizes the effect

of terms that frequently occur in an item and are sig-

niﬁcant for determining the relevance of that item. In

Recommender Systems, a numerical value is assigned

to each term/word in a user proﬁle to indicate the sig-

niﬁcance of each term in deﬁning the user’s interests

(Manning et al., 2009).

In Collaborative Filtering, the system typically

identiﬁes the utility of an item based on user ratings

and ratings from other users for existing items. In

(Adomavicius and Tuzhilin, 2005), it is mentioned

that this allows the system to handle any type of item

(often, there is no representation/characterization of

items, or it is not used in the recommendation pro-

cess, making it unnecessary). The problem arises in

systems where Collaborative Filtering is used, as an

item to be recommended needs to be evaluated by

users, giving rise to the New Item Problem or Item

Cold Start.

One of the most widely used algorithms in Rec-

ommender Systems employing Collaborative Filter-

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

966

ing is the Neighborhood-Based approach, with the

most commonly used variation being User-User if the

system has more items than users. This variation as-

sesses which users have similar tastes (evaluated the

same items and provided similar ratings) compared to

a particular user (Ning et al., 2015).

One of the steps in this approach involves calculat-

ing similarity using the Pearson Coefﬁcient and com-

paring a user with others based on the ratings they

gave to items. The Pearson Coefﬁcient measures the

strength of the relationship between two variables.

Considering this, in the Collaborative Filtering ap-

proach, the user’s proﬁle is represented by a set of

item ratings (Resnick et al., 1994).

When the Recommender System uses a Hybrid

Approach (a mix of Content Based and Collaborative

approaches), the way of representing the user proﬁle

will depend on the kind of item to be recommended.

If an item consists of a textual item (e.g., scientiﬁc pa-

pers), it is possible to use a set of words and a set of

ratings to represent the user proﬁle (Burke, 2002).

In a Demographic Approach, data about the user,

like age, will be necessary. In this case, a pair

attribute-value will represent the user proﬁle. The

problem is that in many contexts, demographic data

will be irrelevant. This is the case of a Recommender

System that recommends scientiﬁc papers, for exam-

ple, (Tintarev and Masthoff, 2015; Adomavicius and

Tuzhilin, 2015). Considering that User Cold Start

is present in more approaches, the emphasis of the

work is on this problem, which involves acquiring

user preferences.

2.2 User Preference Acquisition

Considering the insights from (Knijnenburg and

Willemsen, 2015), four methods can be regarded as

the most commonly used by Recommendation Sys-

tems to acquire these user preferences: item evalu-

ation through scales, assignment of weights to item

attributes, textual reviews of items, and implicit feed-

back.

The method of item evaluation through rating

scales is the most commonly used explicit evaluation

method, where users employ scales (usually a 5-point

Likert scale) to assign ratings to items according to

their preferences (Cena et al., 2010),(Cosley et al.,

2003),(Dooms et al., 2011),(Gena et al., 2011),(Spar-

ling and Sen, 2011). Sometimes, binary evaluations

are made where users are simply asked if they liked

an item or not (Goldberg et al., 2001; Schafer et al.,

2007). On the other hand, assigning weights to item

attributes originates from the ﬁeld of decision anal-

ysis, where multi-attribute utility is used as a stan-

dard for decision-making (Bettman et al., 1998). In

this case, a degree of importance is assigned to each

attribute of the item to be recommended, requiring

the user to indicate the degree of importance for each

item. There is a strong dependence on the type of

item to be recommended, as different classes of items

present different attributes (e.g., for a computer, the

amount of memory, processor, and price are attributes

that may have distinct weights for different users).

The values of these attributes can be discrete values

arranged on a scale indicating from the best value to

the worst value, e.g., for the cost attribute of an item,

the values could be “very high, high, medium, low,

very low” (Bettman et al., 1998).

Another method of eliciting user preferences is

through a critique/comment made for the item. For

example, this comment can even be processed auto-

matically to determine whether the critique was posi-

tive or negative. Finally, considering that users may

not always be willing to provide evaluations, indi-

cate which attributes are most important to them in an

item, or elaborate on comments, many Recommen-

dation Systems seek to observe user behavior, i.e.,

their actions, which does not require extra effort from

the user (Ricci et al., 2015; Amatriain and Basilico,

2015). Thus, actions such as viewing item details,

purchasing/consuming, and selecting items for pur-

chase, among others, indicate interest in the item.

Furthermore, user demographic data such as age,

gender, and occupation can be considered for proﬁle

formation (Adomavicius and Tuzhilin, 2015) and the

creation of stereotypes that are associated with pref-

erences for speciﬁc items. Demographic data signif-

icantly mitigates issues related to User Cold Start.

Still, this data may not always be available, and its

use may not be suitable for recommending certain

items (e.g., recommender systems for scientiﬁc pa-

pers). Some recommender systems utilize user data

obtained from social networks (Son, 2016). However,

the challenge lies in the fact that these data are not al-

ways available. It is important to note that regardless

of the approach used, these ways to acquire informa-

tion can be useful. In this sense, even in the Content-

Based approach, ratings given to items can be utilized.

It is possible, for example, to assign greater weight to

terms in the user’s proﬁle that belong to well-rated

documents.

3 SYSTEMATIC LITERATURE

MAPPING

The Systematic Literature Mapping aims to identify

approaches to minimize User Cold Start presented in

Exploring Strategies to Mitigate Cold Start in Recommender Systems: A Systematic Literature Mapping

967

the literature, excluding approaches that use exter-

nal sources to the Recommender System (e.g., Social

Networks) and considering only the information ob-

tained during the user’s initial access to the system.

The Systematic Literature Mapping was based on (Pe-

tersen et al., 2008).

The expected results for the review include (a)

Identifying techniques/strategies to minimize the

Cold Start for new users in Recommender Systems;

(b) Identifying the approaches of Recommender Sys-

tems (predominantly Content-Based, Collaborative,

or Hybrid) where these techniques/strategies have

been employed; (c) Identifying the application do-

mains of Recommender Systems where these tech-

niques/strategies have been utilized; and (d) Identify-

ing the types of user proﬁle representations employed

in the identiﬁed works.

3.1 Research Questions

Given the objectives, we have formulated a main re-

search question (MRQ) and secondary research ques-

tions (SRQ). The following section presents these

questions: MRQ What techniques have been em-

ployed to reduce the user Cold Start problem? SRQ1:

How is the initial user proﬁle created in these works?

SRQ2: What approaches are used in Recommender

Systems in these works? SRQ3: What are the appli-

cation domains of these works? SRQ4: How has the

evolution of strategies to minimize Cold Start been

over time?

3.2 Search and Selection of Papers

Through an initial search on works related to the re-

search questions, keywords were identiﬁed to form

the search query based initially on two terms: (1)

Recommendation Systems and (2) Cold-Start. The

term User Cold-Start was added as the focus was on

new users. Thus, the works should meet the search

query argument ALL(recommender systems or rec-

ommendation systems) AND ALL(user cold start)

AND ABS(cold start), considering the terms “recom-

mender systems” or “recommendation systems” and

the term “user cold start” anywhere in the paper.

The term “cold start” should be in the paper’s

abstract (“ABS”) to identify papers that address the

Cold Start problem and not just mention the term.

In this sense, it was assumed from some prelimi-

nary searches that papers proposing ways to mini-

mize Cold Start mention the expression “Cold Start”

in their abstract.

Thus, the search strings are deﬁned considering

these criteria and the search engine available in each

digital library. The searches were conducted in Jan-

uary 2023 on the following sources: ACM Digital Li-

brary - https://dl.acm.org/; IEEE - https://www.ieee.

org/ and Scopus - https://www.scopus.com/.

Scopus was considered since its database contains

several other scientiﬁc bases (Buchinger et al., 2014).

The number of papers returned by each engine is:

ACM returned 90 papers, Scopus returned 374 and

IEEE returned 44. Out of the 508 documents ob-

tained, an analysis of information in the attributes of

title, abstract, and keywords was conducted, applying

inclusion and exclusion criteria.

The Inclusion Criteria (IC) used were: IC1: The

publication is written in the English; and IC2: The

publication has more than four pages.Out of the 508

documents returned by the search, 469 papers met the

inclusion criteria.

Subsequently, Exclusion Criteria (EC) were ap-

plied, initially consisting of four criteria: EC1: The

publication is not available for fully open access

through the portal;

; EC2: The publication does

not describe a technique to minimize User Cold Start;

EC3: The publication is not a primary scientiﬁc pa-

per; EC4: The publication is duplicated (i.e., with the

same Digital Object Identiﬁer - DOI).

It was necessary to read the works to apply some

of the exclusion criteria. To expedite the process, this

reading was conducted in three phases: (1) Reading

of the title, abstract, and keywords; (2) Reading of

the introduction and conclusion; and (3) Complete

reading of the paper. If it was impossible to iden-

tify whether the exclusion criterion was met after the

ﬁrst phase, the second phase was carried out. When

still insufﬁcient, the third phase was executed. Af-

ter completing this process and considering the four

exclusion criteria, 324 papers remained. However, it

was noted that several works created the user proﬁle

through the analysis of the user’s interaction history

(here, there is no reference to the user’s initial inter-

action, that is, the one made on the ﬁrst access to the

system, but rather interactions recorded after multi-

ple accesses); several works used demographic data;

and several works used external data sources such as

Social Networks from which the data was extracted.

As the objective of this work is to understand how

to create this proﬁle before any prior interaction and

without using external sources to the recommendation

system, that is, at the moment when the user registers,

a ﬁfth exclusion criterion EC5 was deﬁned.

The EC5 deﬁnes: the publication uses the user’s

system access history (not only the initial access), ex-

ternal source (e.g., user’s social networks), or user de-

mographic data as a way to initiate the proﬁle of a new

A free access portal promoted by the government.

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

968

user.

Thus, the EC5 was applied. In the ACM database,

54 papers were identiﬁed that used user history, 9 pa-

pers used demographic data, and 3 papers used some

external source. In the IEEE database, 22 papers used

history, 6 papers used demographic data, and 5 papers

used some external source. Finally, in the SCOPUS

database, 168 papers used user history, 33 papers used

demographic data, and 5 papers used some external

source. After applying CE5 and removing duplicate

papers, only 19 works were selected for analysis (see

Table 1.

Table 1: Number of papers returned per EC.

Exclusion

Criteria

ACM SCOPUS IEEE Total

Total 83 345 41 469

EC 1 83 249 41 373

EC 2 74 231 39 344

EC 3 73 216 39 326

EC 4 73 212 39 324

EC 5 7 6 6 19

3.3 Selected Papers: Solutions to

Minimize the Cold Start (RQ)

As achieved statistical data, papers are distributed

across the search engines used. The majority of stud-

ies were obtained from ACM, around 36.84%. Fi-

nally, IEEE and Scopus contribute with 31.57% each

to the total result of papers per search engine.

Considering the way to create the initial user pro-

ﬁle from the analysis of the works (Creation of the

Initial User Proﬁle (SRQ1)), it is possible to iden-

tify the following strategies: Strategy 1 Requesting

the user to answer questions; Strategy 2 Requesting

the user’s evaluation for a set of items; Strategy 3

Requesting the user to select items; Strategy 4 Re-

questing the user to select/register keywords/tags; and

Strategy 5 Using popular items among older users (or

users considered experts) of the Recommender Sys-

tem.

Strategy 1 consists of asking the users some ques-

tion(s) as soon as they enter the system to identify

their preferences. This is used by (Christakopoulou

et al., 2016), (Walunj et al., 2022), (Okada et al.,

2021). Questions include “What genres of movies do

you enjoy?” or “What places do you enjoy to visit?”.

In the case of Strategy 2, the user is asked to

rate some items at their ﬁrst entry into the sys-

tem. This is used by (Fel

ıcio et al., 2017), (Lazemi

and Ebrahimpour-Komleh, 2017), (He et al., 2020),

(van der Velde et al., 2021)

Strategy 3 is used by (Hristakeva et al., 2017),

(Crawford, 2012), (Fel

ıcio et al., 2016), (Khan et al.,

2021), (Martins et al., 2013), (Fern

andez et al., 2020).

The technique of selecting/registering key-

words/tags (Strategy 4) is used by(Shi et al., 2021),

(Hristakeva et al., 2017).

Finally, using popular items (Strategy 5) involves

items highly rated by active users or called experts.

This strategy is used by (Chao and Guangcai, 2020),

(Lin et al., 2012), (Amatriain et al., 2009), (Zhang

et al., 2020), (Darshna, 2018).

Approaches Used (SRQ2) - In the ACM database,

6 selected papers are Collaborative Filtering, and 1

is Hybrid Approach. In the IEEE database, the three

main recommendation approaches, Collaborative Fil-

tering (CF), Content-Based (CB), and Hybrid (HB),

are found. In the SCOPUS database, the three main

recommendation approaches are also present.

The application Domains (SRQ3) are presented in

the Table 2.

Table 2: Selected papers per application domains.

Paper Recommended

items

(Shi et al., 2021) News

(Christakopoulou et al.,

2016)

Restaurants

(Fel

ıcio et al., 2017) Movies

(Hristakeva et al., 2017) Documents

(Chao and Guangcai, 2020) Movies

(Lin et al., 2012) News

(Amatriain et al., 2009) Movies

(Crawford, 2012) Movies

(Walunj et al., 2022) Hotels, restaurants

e tourist spots

(Okada et al., 2021) Songs

(Fel

ıcio et al., 2016) Art paintings and

clothes

(Zhang et al., 2020) Movies

(Khan et al., 2021) Food recipes

(Lazemi and Ebrahimpour-

Komleh, 2017)

Jokes

(Martins et al., 2013) Videos and songs

(He et al., 2020) Movies

(Darshna, 2018) Songs and movies

(van der Velde et al., 2021) Learning objects

(Fern

andez et al., 2020) Movies

Regarding the Evolution of Solutions to Minimize

Cold Start Over Time (SRQ4), over the years of pub-

lication of the selected papers, the techniques of re-

questing user evaluation for a set of items and ask-

ing the user to select items are present in a signiﬁ-

cant number of years (Fel

ıcio et al., 2017), (Lazemi

and Ebrahimpour-Komleh, 2017), (He et al., 2020),

(van der Velde et al., 2021), (Hristakeva et al., 2017),

Exploring Strategies to Mitigate Cold Start in Recommender Systems: A Systematic Literature Mapping

969

(Crawford, 2012), (Fel

ıcio et al., 2016), (Khan et al.,

2021), (Martins et al., 2013), (Fern

andez et al.,

2020). From 2016 onward, the technique of ask-

ing the user to answer questions or questionnaires

(Christakopoulou et al., 2016), (Walunj et al., 2022),

(Okada et al., 2021) starts to be seen in publications,

and from 2018 onward, the technique of using pop-

ular items becomes present in the papers (Chao and

Guangcai, 2020), (Lin et al., 2012), (Amatriain et al.,

2009), (Zhang et al., 2020), (Darshna, 2018). In 2021,

the technique of asking the user to select or register

keywords/tags were noticed (Shi et al., 2021), (Hris-

takeva et al., 2017). Over the years, there is also

a growing diversity of techniques, with not just one

technique present in all papers published in the same

year.

4 FINAL CONSIDERATIONS

The systematic literature mapping allowed us to iden-

tify strategies to reduce the User Cold Start problem.

Based on the mapping, it is possible to deﬁne fu-

ture works related to the User Cold Start Problem.

Five strategies were identiﬁed: Strategy 1, Request-

ing the user to answer questions; Strategy 2, Request-

ing the user’s evaluation for a set of items; Strategy

3, Requesting the user to select items; Strategy 4, Re-

questing the user to select/register keywords/tags; and

Strategy 5, Using items popular among older users

(or users considered experts) of the Recommendation

System.

These strategies do not involve the user’s previous

interaction with the system, i.e., a user history or the

use of external sources to build the proﬁle (e.g., user

data from social networks) so that the Recommender

System can construct an initial user proﬁle.

Over the years of publication of the selected pa-

pers, the techniques of asking the user to evaluate a

set of items and requesting the user to select items

are present in many papers. From 2016 onwards, the

strategy of requesting the user to respond to ques-

tions/questionnaires started to appear in publications,

and from 2018 onwards, the popular item strategy be-

came prevalent in the papers. Over the years, there

has also been a noticeable increase in the diversity of

strategies, with not just one strategy present in all pa-

pers published in the same year (SRQ3).

Some considerations can be made about these

strategies, aiming for future works. Firstly, many of

these strategies still involve evaluations from other

users, as the new user is not always treated as the ﬁrst

user when using the system. Therefore, strategies that

can be applied to a new user should be considered,

considering that the new user does not require evalua-

tions from other users. There is a research gap regard-

ing works that use these strategies that do not require

any previous user information before the user enters

the system, as well as not needing a user network like

the above strategy of evaluations from other users. It

can also be noted that a limited number of works deal

with the new user problem in Recommender Systems

that use content-based ﬁltering. This fact had been

observed before the mapping through searches in dig-

ital libraries. Few works bring this approach to mini-

mize the Cold Start associated with the identiﬁed gap;

the works found with this approach use strategies of

requesting user responses to questions/questionnaires

or selecting items.

Strategy 1 requires the deﬁnition of a question-

naire. This is not easy because it requires knowledge

about the domain. Besides, Users may not be willing

to spend time ﬁlling out a questionnaire; they might

respond inadequately and even choose not to use the

system. Strategy 5 can be considered suitable only

if the items consist of texts or are described in tex-

tual format. Thus, it is more suitable for the Rec-

ommender System, which uses a Content-based ap-

proach. Besides, deﬁning how the user will give key-

words/tags is necessary. For example: Will the user

write keywords? How many keywords/tags will the

user select? In the last case, how many keywords/tags

will be presented by the user selected, and which key-

words/tags will be shown to the user? A similar issue

is presented in strategy 2 and in strategy 3. How many

items will the user select/evaluate? How many items

will be presented by the user selected, and which

items will be shown to the user?

Show random items, like in (Crawford, 2012), are

not an ideal solution. The problem is that the user

proﬁle can be limited to part of Recommender Sys-

tem items, and the user proﬁle does not reﬂect all

users’ interests. Thus, in future works, a need arises

to apply item diversity before generating recommen-

dations in the user proﬁle formation. In this sense, to

develop an initial proﬁle that enables recommending

items encompassing all user preferences, one possi-

bility would be to present diversiﬁed items to the user,

allowing them to select what they like.

This study aimed to identify, through a System-

atic Literature Mapping, papers addressing the min-

imization of the Cold Start problem without access

to external sources. As a result, 19 papers presenting

strategies for reducing the Cold Start problem were

identiﬁed (MRQ).

The strategies are identiﬁed by requesting that

the user responds to questions/questionnaires, ask-

ing for the user’s evaluation of a set of items, re-

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

970

questing the user to select items, keyword/tag selec-

tion/registration, and ﬁnally, popular item strategies

(SRQ1). Additionally, it was observed that most stud-

ies focus on the movie domain and domains such as

gastronomy, learning objects, humor entertainment

items, videos, restaurants, papers, news, hotels, mu-

sic, and art items (SRQ3). Hence, it is noteworthy to

highlight that, for future works, certain aspects elu-

cidated in the current study can be considered when

formulating a strategy to mitigate the Cold Start chal-

lenge in Recommender Systems of any domain.

ACKNOWLEDGEMENTS

This research is supported by CNPq/MCTI Nº

10/2023 - UNIVERSAL grant n. 402086/2023-

6, CNPq/MCTI/FNDCT Nº 18/2021 grant n.

405973/2021-7 and CAPES - Financing Code 001.

The research by Jos

e Palazzo M. de Oliveira is

partially supported by CNPq grant 306695/2022-7

PQ-SR. The research by Isabela Gasparini is par-

tially supported by CNPq grant 302959/2023-8 and

FAPESC Edital nº 48/2022 TO n°2023TR000245.

REFERENCES

Abdullah, N. A., Rasheed, R. A., Nizam, M. H., and Rah-

man, M. M. (2021). Eliciting auxiliary information

for cold start user recommendation: A survey. Ap-

plied Sciences (Switzerland), 11(20).

Adomavicius, G. and Tuzhilin, A. (2005). Toward the

next generation of recommender systems: A sur-

vey of the state-of-the-art and possible extensions.

IEEE transactions on knowledge and data engineer-

ing, 17(6):734–749.

Adomavicius, G. and Tuzhilin, A. (2015). Context-Aware

Recommender Systems, pages 191–226. Springer US,

Boston, MA.

Alyari, F. and Jafari Navimipour, N. (2018). Recommender

systems: A systematic review of the state of the art

literature and suggestions for future research. Kyber-

netes, 47(5):985–1017.

Amatriain, X. and Basilico, J. D. (2015). Recommender

systems in industry: A netﬂix case study. In Recom-

mender Systems Handbook.

Amatriain, X., Lathia, N., Pujol, J. M., Kwak, H., and

Oliver, N. (2009). The wisdom of the few: A col-

laborative ﬁltering approach based on expert opin-

ions from the web. Proceedings - 32nd Annual In-

ternational ACM SIGIR Conference on Research and

Development in Information Retrieval, SIGIR 2009,

pages 532–539.

Bettman, J. R., Luce, M. F., and Payne, J. W. (1998).

Constructive Consumer Choice Processes. Journal of

Consumer Research, 25(3):187–217.

Buchinger, D., Cavalcanti, G. A. d. S., and Hounsell,

M. D. S. (2014). Mecanismos de busca acad

emica:

uma an

alise quantitativa. Revista Brasileira de

Computac¸

ao Aplicada, 6(1):108–120.

Burke, R. (2002). Hybrid Recommender Systems : Survey

and Experiments. User Modeling and User-Adapted

Interaction, 12(6):331–370.

Cena, F., Vernero, F., and Gena, C. (2010). Towards

a Customization of Rating Scales in Adaptive Sys-

tems. In De Bra, P., Kobsa, A., and Chin, D., edi-

tors, User Modeling, Adaptation, and Personalization,

pages 369–374, Berlin, Heidelberg. Springer Berlin

Heidelberg.

Chao, X. and Guangcai, C. (2020). Collaborative Filtering

and Leaders’ Advice Based Recommendation System

for Cold Start Users. In Proceedings of the 2020 6th

International Conference on Computing and Artiﬁcial

Intelligence, ICCAI ’20, pages 158–164, New York,

NY, USA. Association for Computing Machinery.

Christakopoulou, K., Radlinski, F., and Hofmann, K.

(2016). Towards conversational recommender sys-

tems. Proceedings of the ACM SIGKDD International

Conference on Knowledge Discovery and Data Min-

ing, 13-17-August-2016(3):815–824.

Cosley, D., Lam, S. K., Albert, I., Konstan, J. A., and Riedl,

J. (2003). Is seeing believing? how recommender sys-

tem interfaces affect users’ opinions. In Proceedings

of the SIGCHI Conference on Human Factors in Com-

puting Systems, CHI ’03, page 585–592, New York,

NY, USA. Association for Computing Machinery.

Crawford, D. (2012). Initial Proﬁle Generation in Recom-

mender Systems Using Pairwise Comparison. Com-

munications of the ACM, 36(11):18.

Darshna, P. (2018). Music recommendation based on

content and collaborative approach & reducing cold

start problem. Proceedings of the 2nd International

Conference on Inventive Systems and Control, ICISC

2018, (Icisc):1033–1037.

Dooms, S., De Pessemier, T., and Martens, L. (2011). An

online evaluation of explicit feedback mechanisms for

recommender systems. pages 391–394.

Fel

ıcio, C. Z., Paix

ao, K. V., Barcelos, C. A., and Preux,

P. (2017). A multi-Armed bandit model selection for

cold-start user recommendation. UMAP 2017 - Pro-

ceedings of the 25th Conference on User Modeling,

Adaptation and Personalization, pages 32–40.

Fel

ıcio, C. Z., De Almeida, C. M., Alves, G., Pereira, F. S.,

Paix

ao, K. V., De Amo, S., and Barcelos, C. A. (2016).

Vp-rec: A hybrid image recommender using visual

perception network. pages 70–77.

Fern

andez, D., Formoso, V., Cacheda, F., and Carneiro, V.

(2020). A Content-Based Approach to Proﬁle Expan-

sion. International Journal of Uncertainty, Fuzziness

and Knowlege-Based Systems, 28(6):981–1002.

Gena, C., Brogi, R., Cena, F., and Vernero, F. (2011). The

Impact of Rating Scales on User’s Rating Behavior. In

Konstan, J. A., Conejo, R., Marzo, J. L., and Oliver,

N., editors, User Modeling, Adaption and Personal-

ization, pages 123–134, Berlin, Heidelberg. Springer

Berlin Heidelberg.

Exploring Strategies to Mitigate Cold Start in Recommender Systems: A Systematic Literature Mapping

971

Goldberg, K., Roeder, T., Gupta, D., and Perkins, C. (2001).

Eigentaste: A Constant Time Collaborative Filtering

Algorithm. Information Retrieval, 4(2):133–151.

He, D. P., He, Z. L., and Liu, C. (2020). Recommendation

algorithm combining tag data and naive bayes classiﬁ-

cation. Proceedings - 2020 3rd International Confer-

ence on Electron Device and Mechanical Engineer-

ing, ICEDME 2020, pages 662–666.

Hristakeva, M., Knoth, P., Kershaw, D., Pettit, B., Jack,

K., Rossetti, M., and Vargas, S. (2017). Build-

ing Recommender systems for scholarly information.

ACM International Conference Proceeding Series,

Part F127853:25–32.

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

(2010). Recommender Systems: An Introduction.

Cambridge University Press.

Khan, M. A., Smyth, B., and Coyle, D. (2021). Address-

ing the complexity of personalized, context-aware and

health-aware food recommendations: an ensemble

topic modelling based approach. Journal of Intelli-

gent Information Systems, 57(2):229–269.

Knijnenburg, B. and Willemsen, M. (2015). Evaluating

Recommender Systems with User Experiments, pages

309–352.

Lazemi, S. and Ebrahimpour-Komleh, H. (2017). Improv-

ing collaborative recommender systems via emotional

features. Application of Information and Communica-

tion Technologies, AICT 2016 - Conference Proceed-

ings, pages 1–5.

Lika, B., Kolomvatsos, K., and Hadjiefthymiades, S.

(2014). Facing the cold start problem in recom-

mender systems. Expert systems with applications,

41(4):2065–2073.

Lin, C., Xie, R., Li, L., Huang, Z., and Li, T. (2012).

PRemiSE: Personalized News Recommendation via

Implicit Social Experts. In Proceedings of the 21st

ACM International Conference on Information and

Knowledge Management, CIKM ’12, pages 1607–

1611, New York, NY, USA. Association for Comput-

ing Machinery.

Manning, C. D., Raghavan, P., and Sch

utze, H. (2009). An

Introduction to Information RetrievalAn Introduction

to Information Retrieval. Number c. Cambridge Uni-

versity Press.

Martins, E. F., Bel

em, F. M., Almeida, J. M., and

Gonc¸alves, M. (2013). Measuring and addressing the

impact of cold start on associative tag recommenders.

WebMedia 2013 - Proceedings of the 19th Brazilian

Symposium on Multimedia and the Web, pages 325–

332.

Monti, D., Rizzo, G., and Morisio, M. (2021). A systematic

literature review of multicriteria recommender sys-

tems, volume 54. Springer Netherlands.

Ning, X., Desrosiers, C., and Karypis, G. (2015). A

Comprehensive Survey of Neighborhood-Based Rec-

ommendation Methods, pages 37–76. Springer US,

Boston, MA.

Okada, K., Tan, P. X., and Kamioka, E. (2021). Five-Factor

Musical Preference Prediction for Solving New User

Cold-Start Problem in Content-Based Music Recom-

mender System. IISA 2021 - 12th International Con-

ference on Information, Intelligence, Systems and Ap-

plications.

Panda, D. K. and Ray, S. (2022). Approaches and algo-

rithms to mitigate cold start problems in recommender

systems: a systematic literature review. Journal of In-

telligent Information Systems, pages 341–366.

Pazzani, M. J. and Billsus, D. (2007). Content-based rec-

ommendation systems. In The adaptive web: methods

and strategies of web personalization, pages 325–341.

Springer.

Petersen, K., Feldt, R., Mujtaba, S., and Mattsson, M.

(2008). Systematic mapping studies in software engi-

neering. 12th International Conference on Evaluation

and Assessment in Software Engineering, EASE 2008,

(February 2015).

Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P., and

Riedl, J. (1994). Grouplens: An open architecture for

collaborative ﬁltering of netnews. In Proceedings of

the 1994 ACM conference on Computer supported co-

operative work, pages 175–186.

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

mender systems: introduction and challenges. Rec-

ommender systems handbook, pages 1–34.

Schafer, J. B., Frankowski, D., Herlocker, J., and Sen, S.

(2007). Collaborative Filtering Recommender Sys-

tems, page 291–324. Springer-Verlag, Berlin, Heidel-

berg.

Shi, S., Ma, W., Wang, Z., Zhang, M., Fang, K., Xu, J.,

Liu, Y., and Ma, S. (2021). WG4Rec: Modeling Tex-

tual Content with Word Graph for News Recommen-

dation. In Proceedings of the 30th ACM International

Conference on Information & Knowledge Manage-

ment, CIKM ’21, pages 1651–1660, New York, NY,

USA. Association for Computing Machinery.

Son, L. H. (2016). Dealing with the new user cold-start

problem in recommender systems: A comparative re-

view. Information Systems, 58:87–104.

Sparling, E. I. and Sen, S. (2011). Rating: How difﬁcult is

it? In Proceedings of the Fifth ACM Conference on

Recommender Systems, RecSys ’11, page 149–156,

New York, NY, USA. Association for Computing Ma-

chinery.

Tintarev, N. and Masthoff, J. (2015). Explaining recom-

mendations: Design and evaluation. In Recommender

systems handbook, pages 353–382. Springer.

van der Velde, M., Sense, F., Borst, J., and van Rijn, H.

(2021). Alleviating the Cold Start Problem in Adap-

tive Learning using Data-Driven Difﬁculty Estimates.

Computational Brain and Behavior, 4(2):231–249.

Walunj, V., Sharma, S., Wagh, A., Solanki, U., and Ma-

hajan, J. (2022). Smart tour advisor using machine

learning and natural language processing. pages 53–

57.

Zhang, Z., Dong, M., Ota, K., and Kudo, Y. (2020). Al-

leviating New User Cold-Start in User-Based Collab-

orative Filtering via Bipartite Network. IEEE Trans-

actions on Computational Social Systems, 7(3):672–

685.

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

972