Content uploaded by Mehdi Elahi
Author content
All content in this area was uploaded by Mehdi Elahi on Apr 18, 2019
Content may be subject to copyright.
“bigdata˙rs˙book2018”
2018/11/16
page 1
Chapter 1
Cold Start Solutions For Recommendation
Systems
Farshad Bakhshandegan Moghaddam1Mehdi Elahi2
This is a preprint of a book chapter accepted by the Book “Big Data Recommender Systems:
Recent Trends and Advances” and is subject to Institution of Engineering and Technology
Copyright. When the final version is published, the copy of record will be available at the IET
Digital Library
Abstract
Recommendation systems are essential tools to overcome the choice overload
problem by suggesting items of interest to users. However, they suffer from a major
challenge which is the so-called cold-start problem. The cold-start problem typi-
cally happens when the system does not have any form of data on new users and on
new items. In this chapter, we describe the cold start problem in recommendation
systems. We mainly focus on Collaborative Filtering (CF) systems which are the
most popular approaches to build recommender systems and have been successfully
employed in many real-world applications. Moreover, we discuss multiple scenarios
that cold-start may happen in these systems and explain different solutions for them.
1.1 Introduction
One of the challenges in everyday life is to make the right decision when purchasing
a product. This challenge has been worsen due to the growing Volume, Variety,
and Velocity of data associated with products3. Although the massive increase in
the number of choices has been an opportunity for consumers to choose the most
interesting products, however, this has led to the problem of Choice Overload, i.e.,
the problem of having unlimited number of choices, especially when they do not
differ significantly from each other [1, 2].
Recommender Systems (RSs) can mitigate this problem by choosing and sug-
gesting a short list of items for users, based on their personal needs and constraints
[3, 4, 5, 6]. These systems, that have been primarily developed and integrated into
1Karlsruhe Institute of Technology
2Free University of Bozen - Bolzano
3https://www.zdnet.com/article/volume-velocity-and-variety- understanding-the- three-vs- of-big- data/
“bigdata˙rs˙book2018”
2018/11/16
page 2
2This is a preprint of a bookchapter in the Book “Big Data Recommender Systems”
the eCommerce websites, have shown to be effective in supporting users when mak-
ing decision. However, their application have gone far beyond that as now they have
been extensively exploited almost anywhere, from social networks to intelligent per-
sonal assistants. Their effectiveness has been proved whenever an enhanced decision
support is required in assisting users during their interaction with a system. Such an
enhanced support enables the users to expand their experience, e.g., by receiving
serendipitous suggestions from a less-explored part of item catalog and allowing the
users to experience surprising items that might not be known to them.
For that aim, recommender systems carefully observe the users’ behaviors and
collect different forms of user preferences, in order to understand the personal tastes
of users [7, 8, 9]. These systems then attempt to filter a long list of items and choose
a shortlist of suggestions. This capability has made them to become an essential
component of any type of commercial information systems that needs to deal with a
large catalog of items [10].
1.1.1 Recommendation Approaches
From the mid-90s when early works on recommender systems [11, 12] have been
emerged, till now, variety of recommendation approaches have been proposed. These
various approaches still share commonalities, based on their underlying algorithms,
that makes it possible to classify them into a number of classes [10, 13, 9, 14]. We
can briefly describe each of these classes based on the definitions in the literature.
One of the most popular class of recommender systems is Collaborative Filter-
ing (CF) [15, 16] which analyzes a set of known ratings and predicts the unknown
ratings, expected to be given to the items by the users. A collaborative filtering
system, then, recommends to a user the items with the highest predicted ratings.
Content-based (CB) [17, 18] class of recommender systems analyze the content of
the items and recommends items based on their associated content attributes (fea-
tures). Utility-based [19, 20] class of recommender systems predicts the utility
scores of users corresponding to the different items (as choice options). This is done
by taking into account the needs and constrains of each user when computing the
utility scores. The items with the highest predicted utility scores are recommended
to the users. Demographic [21, 22] class of recommender systems considers the
demographic data associated with the users and builds recommendations by taking
into account the particular demographic group a user may belong to. Knowledge-
based [23, 24] class of recommender systems adopts a specific reasoning process
which begins by formulating the users’ needs and preferences and ends with identi-
fying whether or not an item matches the specific criteria for a target user. Hybrid
[25, 26] class of recommender systems combines a number of different approaches
from a single or multiple class(es) of recommender systems in order to cope with the
limitations of each single approach.
Regardless of the class of the implemented recommendation approach, a prereq-
uisite to any recommender system is the availability of the data that may indicate the
needs and preferences of the users. Indeed, in spite of the fact that the algorithm per-
formance plays an important role, however, the quality of recommendations based
on any class of recommender systems may become poor if no or low quality data has
“bigdata˙rs˙book2018”
2018/11/16
page 3
Final version will be available at the IET Digital Library 3
Figure 1.1 Rating Matrix: rows represent users and columns represent items. The
entries of the matrix contain the “ known” ratings, users have provided
to items. The “unknown” ratings are represented with question marks.
been provided by users [27, 28]. This is a situation known as Cold Start problem,
which typically happens when a new user registers to the system and no preference
data is available for that user. This is a major problem in recommender systems
specially with large number of users.
In this book chapter, we address the cold start problem in recommender system.
We mainly focus on Collaborative Filtering (CF) systems as they are very popular
type of the real-world recommender systems. We describe different scenarios that
cold start may happen in these systems and survey the solutions for the problem that
have been proposed by the literature.
1.2 Collaborative Filtering
Collaborative filtering based recommender systems exploit a dataset of user feed-
backs, mainly in the form of ratings, that have been provided by a network of users
to a catalog of items. The dataset is typically represented as matrix where rows
represent users and columns represent items (see figure 1.1). Collaborative filter-
ing systems then use this dataset and predict which items could be interesting to a
target user [15, 16]. For that, these systems mine patterns of relationships and sim-
ilarities among the users and use them to learn predictive models that can generate
recommendations.
Such predictions are computed for every unknown rating for a pair of user-item
within the rating matrix. This results in a rank list of items, computed for a target
user, where items are sorted accordingly to their predicted ratings. Collaborative
filtering system selects a short list of items with the highest predicted ratings and
recommends it to the target user.
While recommender systems based on collaborative filtering approach have pre-
sented promising performance, however, they can largely suffer from cold start prob-
lem due to the lack of data for certain users or certain items [13, 29]. The main form
of cold start problem is the New User problem which occurs when a new user regis-
ters to the system and requests to receive recommendations before she has provided
“bigdata˙rs˙book2018”
2018/11/16
page 4
4This is a preprint of a bookchapter in the Book “Big Data Recommender Systems”
Figure 1.2 Illustration of Cold Start problem in recommender systems: New user
problem (left) and New item problem (right)
any rating to any item (see figure 1.2). Another type of cold start is the New Item
problem which occurs when a new item is added to the item catalog and none of the
users has yet rated that new item (see figure 1.2). The Sparsity of the data can be
also considered as relevant issue to the cold start problem. In severe cases of data
sparsity, the performance of the collaborative filtering systems can be seriously dam-
aged leading to a very poor quality of recommendation. This is a situation where the
number of known ratings is extremely smaller than the number of unknown ratings
and the system has to compute predictions for the unknown ratings [13, 30].
The remaining sections, discuss a set of solutions, that have been proposed by
the literature, in addressing the cold start problem.
1.3 Active Learning in Recommender Systems
One of the main solutions to the cold start problem in recommender systems, is
Active Learning. Generally, active learning is part of a broader research topic of
Machine Learning, a well-known research area which focuses on design and devel-
opment of novel algorithms in solving a large variety of tasks such as regression and
classification tasks [31, 32, 33, 34]. These algorithms typically need big datasets
to learn patterns behind data and build models that can be used to predict unprece-
dented data [35]. This is a form of learning process that is called Passive Learning
[36]. However, the availability of such big data can not be always presumed as
there are realistic cases where the data is (e.g.,) partially available. In such cases,
the system may not be able to achieve a certain level of accuracy unless more data
is collected. While this could be beneficial, however, collecting more data can be
an expensive process and may require extensive human involvement. Therefor, the
system has to focus on collecting only high quality data by carefully controlling the
data collection process. This will help the system to minimize the cost of data collec-
“bigdata˙rs˙book2018”
2018/11/16
page 5
Final version will be available at the IET Digital Library 5
Figure 1.3 Active Learning in recommender systems: a new user registers to the
system, where active learner proposes her to rate a selection of items,
and elicits the ratings.
tion while maximizing the expected benefit. This form of learning process is called
Active Learning [36, 37, 28].
In comparison to the passive learning, there can be two big advantageous, brought
by the active learning. The first advantage is that, in active learning, the system does
not need to have access to the entire data and instead, it can iteratively obtain further
data. The second advantage is that, active learning allows the system to carefully
analyze the available data and decide which data to be collected. This process will
disallow the noisy data to be collected and may improve the quality of the input data.
In the context of recommender systems, active learning can bring similar advan-
tageous, and hence, it can be a natural solution to the cold start problem. This can be
the reason why the initial interaction of new users with recommender systems begins
with active learning where the system requests the new users to provide ratings for
a set of selected items [38]. This allows the system to obtain a minimum amount
of data that can describe the preference of a new user (see figure 1.3). Hence, an
active learner follows a set of defined rules that is used to automatically regulate the
item selection process. By applying these rules, the system elicits ratings from the
(new) users and use them to build or update their profiles. The very precise defini-
tion for selecting items to propose to a user to rate is called active learning Strategy
[8, 39, 40, 28].
While there has been a broad range of active learning strategies, proposed in the
literature, however, according to [38, 28], these strategies can be classified into few
classes, listed below:
•Uncertainty Reduction [8, 39] strategies try to select items with more diverse
ratings as the system is less certain about them. Suppose that a lot of users have
given high ratings to an item, while many users have given low ratings to the
same item. In such a case, it will be difficult for the system to predict whether
“bigdata˙rs˙book2018”
2018/11/16
page 6
6This is a preprint of a bookchapter in the Book “Big Data Recommender Systems”
or not to recommend that item. Conversely, an item that has received low ratings
from nearly all users can be easily excluded from the recommendation. Hence,
collecting the ratings of items with diverse ratings may be very informative and
may decrease the uncertainty of the system when computing predictions [8, 39].
•Error Reduction [40, 28] strategies attempt to select items that collecting their
ratings may directly reduce the prediction error. This is due to the fact that there
are items with highly diverse ratings where the ratings are poorly correlated with
the ratings of the other items (e.g., Napoleon Dynamite movie in Netflix dataset)
[40]. While selection of such items for active learning may not contribute to
the predictive power of the system, still uncertainty reduction strategies may
select them for active learning. Instead, error reduction strategies may ignore
these items and focus more on items with ratings that can positively improve
the prediction accuracy [28].
•User Adaptation [41, 39] strategies try to personalize the active learning pro-
cess to the particular characteristics of the users by selecting and proposing
different items for different users to rate. This is due to the fact that different
users may have different knowledge, familiarity and preferences towards dif-
ferent category of items and hence it is not very convenient to select a similar
set of items for these different types of users. Accordingly, taking into account
such differences among users in active learning process could lead to collecting
higher quality and quantity of ratings.
•Acquisition Probability [28, 42] strategies try to maximize the chance that a
user can rate an item and hence they select items that are more likely to be
known by a user. Suppose that a user has not been in a restaurant while the
system requests her to rate that restaurant. The rating of that user may not be so
informative and instead may increase the level of noise within the data. Hence,
it is crucial for the active learner to take into account the likelihood that a user
is familiar with an item when requesting her to rate the item.
•Decision Tree-based [39, 43] strategies adopt decision tree algorithms in order
to identify informative items to be selected for active learning. Each node of
such decision tree, contains a candidate item to be proposed to a new user to
rate. Therefore, the node somehow represents a group of like-minded users who
has rated that candidate item similarly. Accordingly, each node splits the users
into 3 groups, i.e., those who have given that candidate item (i) high rating,
(ii) low ratings, or (iii) no rating. The active learner builds this decision tree
based on an optimization term that leads to the reduction of the prediction error.
Once the decision tree is built, the system can use it to iteratively select items
to propose to a new user, hence traversing from root node of the tree to the leaf
nodes, depending on the ratings provided by the user.
•Prediction based [37, 44] strategies build prediction models that are used to
decide which items to be selected for active learning. The prediction-based
strategies rank items according to the predicted ratings and select the top items
with highest predicted ratings. The adopted predictive models may vary from
Probabilistic models [45, 41, 28] to Matrix Factorization models [46, 47]. An
advantageous of these strategies is that they select items that are likely to be in-
“bigdata˙rs˙book2018”
2018/11/16
page 7
Final version will be available at the IET Digital Library 7
teresting for users and hence the users are not bothered during the active learn-
ing process. Indeed, the users may even enjoy checking and rating the selected
items. It is also highly probable that the proposed items are familiar to the users,
and hence, the chance to actually obtain the ratings by these strategies is high.
•Hybrid [48, 49] strategies combine a number of individual strategies in order
to take advantages of multiple ones. This may allow the hybrid strategies to
simultaneously optimize different metrics, such as accuracy,diversity, and user
satisfaction. Moreover, there are situations that an individual strategy may fail
to properly select items to propose to a target user to rate. However, in such
particular situations, hybridizing the individual strategies can tackle the problem
and lead to improving the performance of the individual strategies.
1.4 Semantic-based Recommender Systems
The traditional solutions for the cold start problem are based on the popular Content-
based Filtering (CBF) approaches. These approaches build user profiles by associ-
ating their preferences with the semantic attributes of the item content [50, 51, 52,
6, 53, 54]. Exploiting the content of the items has been used to address the new
item problem. When a new item is added to the catalog, the item profile is built by
various types of semantic attributes (see figure 1.4). The recommender can use such
profiles to compute similarity or built machine learning models to generate relevant
recommendations.
In early recommender systems, semantic attributes were based on less-structured
form of semantic content such as item category or item description. These attributes
are exploited by the recommender systems to establish Vector Space Model (VSM)
[18], where, each item is represented by a multi-dimensional vector of content at-
tributes [55].
More novel class recommender systems has been emerged after the famous ar-
ticle of Tim Berners-Lee 4(as known as the father of the Semantic Web) [56]. He
proposed to formulate a set of rules to create the Web of data, known as Linked
Data principles [57]. In order to better understand Linked Data, the following brief
description of content architecture in Web could be beneficial.
Current Web, as known as Web of Document, contains billions of documents
which are related to each other by hyperlinks. This architecture makes it possible
for users to traverse the Web by visiting hyperlinks. While the content of the Web
is human-readable, however, it still suffers from massive ambiguity originated from
the lack of a proper structure with respect to the representation of information. This
ambiguity in information consequently makes it incapable for machines to under-
stand the provided information. Linked Data principles [57] are indeed proposed in
addressing this problem.
According to the noted proposal, the knowledge is modeled by Resource De-
scription Framework (RDF) which provides a generic graph-based data model for
describing resources, including their relationships with other resources [58]. By
4https://en.wikipedia.org/wiki/Tim Berners-Lee
“bigdata˙rs˙book2018”
2018/11/16
page 8
8This is a preprint of a bookchapter in the Book “Big Data Recommender Systems”
Figure 1.4 Recommendation based on semantic attributes, in addressing the New
Item problem as part of Cold Start problem
interlinking some publicly available linked data dataset such as DBpedia5and Wiki-
data6,Linked Open Data has been emerged as a network of interconnected datasets,
accessible via endpoints. It is possible to query Linked Open Data by query lan-
guages such as RDF Query Language [59].
Such database has been used by novel class of recommender systems that are
relied on the new form of semantic data that can better represent the knowledge
of human. These novel semantic recommender systems focused on exploiting the
semantic content information rather than the raw content data based on the Web of
data [60]. This has brought variety of advantages to recommender systems, such as,
mitigating the new item cold start problem, as well as, empowering recommender
systems to provide semantic-aware explanations for recommendations.
1.5 Recommendation based on Visual Features
Another group of recommendation approaches, that can address the cold-start prob-
lem, implements the idea of enriching the item profiles with additional source of
data. The enriching mechanism allows them to be capable of coping with the New
Item problem. A representative technique within this group of recommender sys-
tems is proposed by [61] where the authors exploited a set of visual features in a
multimedia recommender system (see figure 1.5). The proposed features are called
5https://wiki.dbpedia.org/
6https://www.wikidata.org/wiki/Wikidata:Main Page
“bigdata˙rs˙book2018”
2018/11/16
page 9
Final version will be available at the IET Digital Library 9
Figure 1.5 Recommendation based on visual features, in addressing the New Item
problem as part of Cold Start problem
Mise-en-scene features and they are based on variation of colours, camera and object
motions, and lighting within the multimedia items. The results of the experiments
have shown that these automatically extracted features can solve the new item cold
start problem [62, 61, 63]. The same authors have extended that work and proposed
another recommendation technique based on exploiting MPEG7 and Deep Learning
visual features [64]. Again, the results have shown the substantial power of visual
features in solving the new item problem in recommender systems. There have been
many recent related works that have used visual features in recommender systems,
but mainly focused on deep learning features [65, 66, 67, 68].
There have been also earlier works that have studied the potential of building
style-aware recommender systems based on visual features [69, 70, 71, 72, 73, 74].
As an example, the authors of [72] introduced VideoReach which is a recommender
system that can extend the semantic item profile with visual features. The results of
their experiments have shown that this extension has positively affected the click-
through-rate. The work in [73] presented an algorithm that can integrates different
ranking lists, generated based on visual features and none-visual attributes. The
results have shown improvement.
A limitation of these works is that they have typically assumed that there are
already a set of semantic attributes collected and the visual features are used in com-
bination with these semantic attributes. Therefore, further studies is needed in in-
vestigating the actual power of visual features, mainly when traditional semantic
attributes are not available.
“bigdata˙rs˙book2018”
2018/11/16
page 10
10 This is a preprint of a bookchapter in the Book “Big Data Recommender Systems”
Figure 1.6 Personality based recommender systems, in addressing the New User
problem as part of Cold Start problem
1.6 Personality-based Recommender Systems
One of natural solutions to tackle with the cold start problem is to use additional
user attributes (as known as side information), in order to build the initial profile of
a new user [75, 76]. There have been already different types of such attributes, pro-
posed in the literature. However, one of the most representative forms of attributes
is the psychological ones related to the Personality Traits of users. These personal
traits are based on predictable and stable characteristics of users, and they describe
the “consistent behavior pattern and interpersonal processes originating within the
individuals” [77]. Personality traits can portray the differences of users in terms of
emotional, interpersonal, experiential, attitudinal and motivational aspects [78].
Psychology literature is already mature in the personality field and various psy-
chological models are available on how to represent the personality aspects of an
individual person. One of the most well-known models is the Five Factor model
(FFM) [79], which is commonly adopted in different research disciplines [80]. This
model describes the personality of a person with respect to five dimensions as known
as Big Five traits: Openness, Conscientiousness, Extroversion, Agreeableness and
Neuroticism (as known as OCEAN).
It has been shown that users with different personality traits express differences
in their decision making process [81, 82]. Accordingly, users with similar personality
traits are more likely to share similar preferences [83]. Authors of [84] have studied
the correlation of personality traits with musical preferences and showed that users
with high openness trait typically share similar preferences for jazz, blues and clas-
sical musical genres, and users with high extroversion and agreeableness traits are
likely to enjoy rap, hip-hop, funk and electronic musical genres. The authors of [85]
have conducted an experiment that showed a strong relation between the preferences
of users for certain web applications and their particular personality traits. In [86],
“bigdata˙rs˙book2018”
2018/11/16
page 11
Final version will be available at the IET Digital Library 11
the relation of personality traits and emotional expressions have been investigated
for users who were watching movies in different social contexts. The results have
showed that different patterns of emotional expressions can be observed for different
users with their unique personality traits.
The promising results of the above-described works, showing the correlation
of personality and preferences of users, has motivated further studies on the idea
of exploiting personality in recommender systems, e.g., in addressing the cold start
scenario [44, 30, 87, 82, 83]. Hence, when a new user enters the system and has not
provided any data associated with her preferences, personality traits can be used to
profile her and generate personalized recommendation (see figure 1.6). Hence, the
personality can be used either to compute the similarity among users for similarity-
based recommender systems, or as additional user attributes, in model based recom-
mender systems.
As an example of works within this area, the authors of [88] adopted different
recommendation approaches and showed that incorporation of personality may lead
to a better recommendation quality in cold-start scenario. [89] has investigated the
potential of using personality and showed that personality characteristics can lead to
improvement in the performance of recommender systems. In [90, 80] the relation
of personality of musical tastes is exploited in order to generate relevant recom-
mendation for users. Finally, [91] has developed a recommender system that uses
personality profiles of users to generate recommendations for them. This is done by
first analyzing the hotel reviews written by users. Then using the correlations among
the reviews and the personality traits, the system extracts the personality profiles of
the users and compute similarities among the users in order to build similarity-based
recommendations.
A limitation of personality-based approaches is that, before the personality data
is used, the users should complete a personality questionnaire, which can be a time
consuming process. This is why there are recent machine learning techniques that
are built to extract the personality traits from other sources such as social network
profiles of the users.
1.7 Cross Domain Recommender Systems
Another solution for the cold start problem in recommender systems is based on
exploitation of axillary domains in order to generate recommendations in a target
domain. This is called Cross-domain Recommendation and it is one of the research
topics that have been well-studied in the community of recommender systems. The
reason can be due to the fact that current e-commerce web applications typically
operate in multiple domains and they use mechanisms to aggregate multiple types
of data from multiple domains. Availability of such data can bring benefits to a
recommender system and enables it to perform (e.g.,) cross-selling or coping with
the cold start problem in its target domain.
There have been various algorithms developed for cross-domain recommen-
dation [92, 93, 94]. While these algorithms may implement different mechanisms
for the cross-domain recommendation, they share commonalities which enables us
“bigdata˙rs˙book2018”
2018/11/16
page 12
12 This is a preprint of a bookchapter in the Book “Big Data Recommender Systems”
Figure 1.7 Cross-domain recommender systems, in addressing the New User
problem as part of Cold Start problem
to classify them into two major classes, i.e., Knowledge Aggregation approaches
[95, 96, 97, 98] and Knowledge Transfer approaches [99, 100, 101, 102].
The former approach aims to aggregate the knowledge from different auxiliary
domains in order to generate recommendations in the target domain. The latter ap-
proach is based on the idea of eliciting the user ratings from auxiliary domains and
transfer this knowledge to the target domain. In this sense, the latter approach at-
tempts to link different domain knowledges in order to support the recommendation
for the target domain [99].
As an example of former approach can be the work in [103], that proposed var-
ious knowledge aggregation mechanisms that have proved to be effective in improv-
ing the accuracy of target domain recommendations in cold start. An example of the
latter approach is presented in [104] where the authors propose leveraging the prefer-
ence knowledge transfer from an auxiliary domain to the target domain. The results
of evaluation have shown that the proposed recommendation method overtakes the
classical recommendation methods.
A limitation of the cross-domain recommendation is that, there has to be con-
siderable overlap among the adopted datasets in different domains. Hence, without
the having axillary domain and the target domain overlap, it would be not feasible to
apply the techniques described in this section.
1.8 Conclusion
In this book chapter, we addressed the cold start problem in recommender systems.
This problem happens when the system is not able to recommend relevant items to a
new user or to recommend a new item to the existing users.
“bigdata˙rs˙book2018”
2018/11/16
page 13
Final version will be available at the IET Digital Library 13
Table 1.1 Summary of the solutions for the Cold Start problem
Solution Cold start Methods
New User New Item
Active Learning
•Uncertainty Reduction [105, 8, 39]
•Error Reduction [40, 106]
•User Adaptation [107, 40, 8]
•Acquisition Probability [108, 109,
110]
•Decision Tree-based [39, 43, 47]
•Prediction-based [41, 46, 37]
•Hybrid [48, 111, 37]
Semantic Attributes •Graph-based [112, 113]
•Machine Learning [114, 115, 116]
Personality Traits •Similarity based on personality [103]
•Personality traits in the model [104]
Visual Features
•Mise-en-scene features [61, 62, 117]
•MPEG7 features [63]
•Deep Learning features [66, 67, 68]
Cross-domain •Knowledge Aggregation [103]
•Knowledge Transfer [104]
We discussed various solutions that have been proposed in the literature. These
solutions are summarized in table 1.1. These solutions can be classified into 5
classes, i.e., Active Learning,Semantic Attributes,Visual Features,Personality Traits,
and Cross-domain Recommendation. Although all of these solutions have been suc-
cessfully applied and evaluated in prior works, however, none of these solutions can
be seen as a conclusive remedy to the cold start as a generic problem. Indeed, each of
these solutions can be effective in a particular situation of cold start. Some of these
solutions (semantic attributes and visual features) can address the new item problem
while some others (personality traits and cross-domain) can address the new user
problem. Active learning techniques can address both of these problems.
It is worth noting that, the cold start research area in recommender systems is
a multi-disciplinary field of research, and involves disciplines of Machine Learn-
ing, Psychology, and Human Computer Interaction (HCI). For instance, each of the
cold start solutions need proper adoption of the interface design patterns [118] when
obtaining user preferences or presenting a recommended item. Therefore, collabo-
ration among researchers within these disciplines can surely be useful in improving
the quality of current the state-of-the-art approaches.
In conclusion, this chapter shall hopefully provide an overall overview of the
research on cold start and can be a useful source of guidelines for researchers in the
academia and practitioners in the industry. It can hopefully advances the knowledge
in this area, as well as, the related areas.
“bigdata˙rs˙book2018”
2018/11/16
page 14
14 Big Data Recommender Systems: Recent Trends and Advances
References
[1] Bollen D, Knijnenburg BP, Willemsen MC, et al. Understanding choice
overload in recommender systems. In: Proceedings of the fourth ACM con-
ference on Recommender systems. ACM; 2010. p. 63–70.
[2] Anderson C. The Long Tail. Random House Business; 2006.
[3] Resnick P, Varian HR. Recommender systems. Commun ACM.
1997;40(3):56–58.
[4] Shardanand U, Maes P. Social Information Filtering: Algorithms for Au-
tomating &Ldquo;Word of Mouth&Rdquo;. In: Proceedings of the SIGCHI
Conference on Human Factors in Computing Systems. CHI ’95. New York,
NY, USA: ACM Press/Addison-Wesley Publishing Co.; 1995. p. 210–217.
Available from: http://dx.doi.org/10.1145/223904.223931.
[5] Ricci F, Rokach L, Shapira B, et al. Recommender Systems Handbook.
Springer; 2011.
[6] Jannach D, Zanker M, Felfernig A, et al. Recommender Systems: An Intro-
duction. Cambridge University Press; 2010.
[7] Rubens N, Kaplan D, Sugiyama M. Active Learning in Recommender Sys-
tems. In: Ricci F, Rokach L, Shapira B, et al., editors. Recommender Sys-
tems Handbook. Springer Verlag; 2011. p. 735–767.
[8] Rashid AM, Albert I, Cosley D, et al. Getting to Know You: Learning New
User Preferences in Recommender Systems. In: Proceedings of the 2002 In-
ternational Conference on Intelligent User Interfaces, IUI 2002. ACM Press;
2002. p. 127–134.
[9] Su X, Khoshgoftaar TM. A survey of collaborative filtering techniques. Adv
in Artif Intell. 2009 Jan;2009:4:2–4:2. Available from: http://dx.doi.org/10.
1155/2009/421425.
[10] Burke R. Hybrid Recommender Systems: Survey and Experiments. User
Modeling and User-Adapted Interaction. 2002;12(4):331–370. Available
from: ./papers/burke-umuai-ip- 2002.pdf.
[11] Resnick P, Iacovou N, Suchak M, et al. GroupLens: An Open Architecture
for Collaborative Filtering of Netnews. In: Proceedings of the 1994 ACM
Conference on Computer Supported Cooperative Work. CSCW ’94. New
York, NY, USA: ACM; 1994. p. 175–186. Available from: http://doi.acm.
org/10.1145/192844.192905.
[12] Shardanand U, Maes P. Social Information Filtering: Algorithms for Au-
tomating &Ldquo;Word of Mouth&Rdquo;. In: Proceedings of the SIGCHI
Conference on Human Factors in Computing Systems. CHI ’95. New York,
NY, USA: ACM Press/Addison-Wesley Publishing Co.; 1995. p. 210–217.
Available from: http://dx.doi.org/10.1145/223904.223931.
[13] Adomavicius G, Tuzhilin A. Toward the next generation of recommender
systems: a survey of the state-of-the-art and possible extensions. Knowledge
and Data Engineering, IEEE Transactions on. 2005 June;17(6):734–749.
“bigdata˙rs˙book2018”
2018/11/16
page 15
REFERENCES 15
[14] Ricci F, Rokach L, Shapira B. Introduction to recommender systems hand-
book. In: Ricci F, Rokach L, Shapira B, et al., editors. Recommender Sys-
tems Handbook. Springer Verlag; 2011. p. 1–35.
[15] Koren Y, Bell R. Advances in Collaborative Filtering. In: Ricci F, Rokach
L, Shapira B, et al., editors. Recommender Systems Handbook. Springer
Verlag; 2011. p. 145–186.
[16] Desrosiers C, Karypis G. A Comprehensive Survey of Neighborhood-based
Recommendation Methods. In: Ricci F, Rokach L, Shapira B, et al., editors.
Recommender Systems Handbook. Springer; 2011. p. 107–144.
[17] Balabanovi´
c M, Shoham Y. Fab: Content-based, Collaborative Recom-
mendation. Commun ACM. 1997 Mar;40(3):66–72. Available from:
http://doi.acm.org/10.1145/245108.245124.
[18] Pazzani MJ, Billsus D. The Adaptive Web. Berlin, Heidelberg: Springer-
Verlag; 2007. p. 325–341. Available from: http://dl.acm.org/citation.cfm?
id=1768197.1768209.
[19] Guttman RH, Moukas AG, Maes P. Agent-mediated Electronic Commerce:
A Survey. Knowl Eng Rev. 1998 Jul;13(2):147–159. Available from: http:
//dx.doi.org/10.1017/S0269888998002082.
[20] Huang SL. Designing Utility-based Recommender Systems for e-
Commerce: Evaluation of Preference-elicitation Methods. Electron Com-
mer Rec Appl. 2011 Jul;10(4):398–407. Available from: http://dx.doi.org/
10.1016/j.elerap.2010.11.003.
[21] Pazzani MJ. A Framework for Collaborative, Content-Based and Demo-
graphic Filtering. Artif Intell Rev. 1999 Dec;13(5-6):393–408. Available
from: http://dx.doi.org/10.1023/A:1006544522159.
[22] Wang Y, Chan SCF, Ngai G. Applicability of Demographic Recommender
System to Tourist Attractions: A Case Study on Trip Advisor. In: Pro-
ceedings of the The 2012 IEEE/WIC/ACM International Joint Conferences
on Web Intelligence and Intelligent Agent Technology - Volume 03. WI-
IAT ’12. Washington, DC, USA: IEEE Computer Society; 2012. p. 97–101.
Available from: http://dx.doi.org/10.1109/WI-IAT.2012.133.
[23] Burke R. Knowledge-Based Recommender Systems; 2000.
[24] Felfernig A, Burke R. Constraint-based Recommender Systems: Technolo-
gies and Research Issues. In: Proceedings of the 10th International Con-
ference on Electronic Commerce. ICEC ’08. New York, NY, USA: ACM;
2008. p. 3:1–3:10. Available from: http://doi.acm.org/10.1145/1409540.
1409544.
[25] Claypool M, Gokhale A, Miranda T, et al. Combining content-based and
collaborative filters in an online newspaper. In: Proceedings of the ACM
SIGIR ’99 Workshop on Recommender Systems: Algorithms and Evalua-
tion. Berkeley, California: ACM; 1999. .
[26] Li Q, Kim BM. An Approach for Combining Content-based and Collabora-
tive Filters. In: Proceedings of the Sixth International Workshop on Infor-
mation Retrieval with Asian Languages - Volume 11. AsianIR ’03. Strouds-
“bigdata˙rs˙book2018”
2018/11/16
page 16
16 Big Data Recommender Systems: Recent Trends and Advances
burg, PA, USA: Association for Computational Linguistics; 2003. p. 17–24.
Available from: http://dx.doi.org/10.3115/1118935.1118938.
[27] Elahi M, Ricci F, Rubens N. A Survey of Active Learning in Collaborative
Filtering Recommender Systems. Comput Sci Rev. 2016 May;20(C):29–50.
Available from: http://dx.doi.org/10.1016/j.cosrev.2016.05.002.
[28] Rubens N, Elahi M, Sugiyama M, et al. Active Learning in Recommender
Systems. In: Recommender Systems Handbook - chapter 24: Recommend-
ing Active Learning. Springer US; 2015. p. 809–846.
[29] Schein AI, Popescul A, Ungar LH, et al. Methods and metrics for cold-start
recommendations. In: SIGIR ’02: Proceedings of the 25th annual interna-
tional ACM SIGIR conference on Research and development in information
retrieval. New York, NY, USA: ACM; 2002. p. 253–260.
[30] Braunhofer M, Elahi M, Ricci F. Techniques for cold-starting context-
aware mobile recommender systems for tourism. Intelligenza Artificiale.
2014;8(2):129–143.
[31] Sipser M. Introduction to the Theory of Computation. 1st ed. International
Thomson Publishing; 1996.
[32] Alpaydin E. Introduction to Machine Learning. 2nd ed. The MIT Press;
2010.
[33] Abu-Mostafa YS, Magdon-Ismail M, Lin HT. Learning From Data. AML-
Book; 2012.
[34] Flach P. Machine Learning: The Art and Science of Algorithms That Make
Sense of Data. New York, NY, USA: Cambridge University Press; 2012.
[35] Ge M, Helfert M. A Review of Information Quality Research - Develop a
Research Agenda. In: ICIQ; 2007. p. 76–91.
[36] Tong S. Active Learning: Theory and applications. The Department of
Computer Science; 2001.
[37] Elahi M, Ricci F, Rubens N. Active Learning Strategies for Rating Elicita-
tion in Collaborative Filtering: a System-Wide Perspective. ACM Transac-
tions on Intelligent Systems and Technology. 2014;5(1).
[38] Elahi M, Ricci F, Rubens N. Active learning in collaborative filtering recom-
mender systems. In: E-Commerce and Web Technologies. Springer; 2014.
p. 113–124.
[39] Rashid AM, Karypis G, Riedl J. Learning preferences of new users in rec-
ommender systems: an information theoretic approach. SIGKDD Explor
Newsl. 2008 December;10:90–100. Available from: http://doi.acm.org/10.
1145/1540276.1540302.
[40] Golbandi N, Koren Y, Lempel R. On bootstrapping recommender systems.
In: Proceedings of the 19th ACM international conference on Informa-
tion and knowledge management. CIKM ’10. New York, NY, USA: ACM;
2010. p. 1805–1808. Available from: http://doi.acm.org/10.1145/1871437.
1871734.
[41] Harpale AS, Yang Y. Personalized active learning for collaborative filter-
ing. In: SIGIR ’08: Proceedings of the 31st annual international ACM
“bigdata˙rs˙book2018”
2018/11/16
page 17
REFERENCES 17
SIGIR conference on Research and development in information retrieval.
New York, NY, USA: ACM; 2008. p. 91–98.
[42] He L, Liu NN, Yang Q. Active Dual Collaborative Filtering with Both Item
and Attribute Feedback. In: AAAI; 2011. .
[43] Golbandi N, Koren Y, Lempel R. Adaptive bootstrapping of recommender
systems using decision trees. In: Proceedings of the fourth ACM interna-
tional conference on Web search and data mining. WSDM ’11. New York,
NY, USA: ACM; 2011. p. 595–604. Available from: http://doi.acm.org/10.
1145/1935826.1935910.
[44] Elahi M, Braunhofer M, Ricci F, et al. Personality-Based Active Learning
for Collaborative Filtering Recommender Systems. In: Baldoni M, Baroglio
C, Boella G, et al., editors. AI*IA. vol. 8249 of Lecture Notes in Computer
Science. Springer; 2013. p. 360–371. Available from: http://dblp.uni-trier.
de/db/conf/aiia/aiia2013.html#ElahiBRT13.
[45] Jin R, Si L. A Bayesian Approach toward Active Learning for Collaborative
Filtering. In: UAI ’04, Proceedings of the 20th Conference in Uncertainty
in Artificial Intelligence, July 7-11 2004, Banff, Canada; 2004. p. 278–285.
[46] Karimi R, Freudenthaler C, Nanopoulos A, et al. Non-myopic active learn-
ing for recommender systems based on Matrix Factorization. In: IRI. IEEE
Systems, Man, and Cybernetics Society; 2011. p. 299–303.
[47] Karimi R, Freudenthaler C, Nanopoulos A, et al. Active learning for aspect
model in recommender systems. In: CIDM. IEEE; 2011. p. 162–167.
[48] Rubens N, Sugiyama M. Influence-based collaborative active learning. In:
Proceedings of the 2007 ACM conference on Recommender systems. Rec-
Sys ’07. New York, NY, USA: ACM; 2007. p. 145–148. Available from:
http://doi.acm.org/10.1145/1297231.1297257.
[49] Elahi M, Ricci F, Rubens N. Adapting to natural rating acquisition with
combined active learning strategies. In: ISMIS’12: Proceedings of the 20th
international conference on Foundations of Intelligent Systems. Berlin, Hei-
delberg: Springer-Verlag; 2012. p. 254–263.
[50] Degemmis M, Lops P, Semeraro G. A content-collaborative recommender
that exploits WordNet-based user profiles for neighborhood formation. User
Modeling and User-Adapted Interaction. 2007;17(3):217–255.
[51] Eirinaki M, Vazirgiannis M, Varlamis I. SEWeP: using site semantics and a
taxonomy to enhance the Web personalization process. In: Proceedings of
the ninth ACM SIGKDD international conference on Knowledge discovery
and data mining. ACM; 2003. p. 99–108.
[52] Magnini B, Strapparava C. Improving user modelling with content-based
techniques. In: User Modeling 2001. Springer; 2001. p. 74–83.
[53] Garca-Crespo , Chamizo J, Rivera I, et al. SPETA: Social perva-
sive e-Tourism advisor. Telematics and Informatics. 2009;26(3):306–
315. Available from: http://dblp.uni-trier.de/db/journals/tele/tele26.html#
Garcia-CrespoCRMPB09.
“bigdata˙rs˙book2018”
2018/11/16
page 18
18 Big Data Recommender Systems: Recent Trends and Advances
[54] Towle B, Quinn C. Knowledge Based Recommender Systems Using Ex-
plicit User Models. In: Papers from the AAAI Workshop, AAAI Technical
Report WS-00-04. Menlo Park, CA: AAAI Press; 2000. p. 74–77.
[55] Lops P, De Gemmis M, Semeraro G. Content-based recommender systems:
State of the art and trends. In: Recommender systems handbook. Springer;
2011. p. 73–105.
[56] Berners-Lee T, Hendler J, Lassila O. The Semantic Web. Scientific Amer-
ican. 2001 May;284(5):34–43. Available from: http://www.sciam.com/
article.cfm?articleID=00048144-10D2-1C70-84A9809EC588EF21.
[57] Berners-Lee T. Linked-data design issues; 2009.
Http://www.w3.org/DesignIssue/LinkedData.html. W3C design issue doc-
ument. Available from: http://www.w3.org/DesignIssues/LinkedData.html.
[58] Hitzler P, Krtzsch M, Rudolph S. Foundations of Semantic Web Technolo-
gies. CRC Press; 2010.
[59] Krummenacher R, Norton B, Marte A. Towards Linked Open Services and
Processes. In: Berre AJ, G´
omez-P´
erez A, Tutschku K, et al., editors. Future
Internet - FIS 2010 - Proceedings of the Third Future Internet Symposium,
Berlin, Germany, September 20-22, 2010. vol. 6369 of Lecture Notes in
Computer Science. Springer; 2010. p. 68–77. Available from: http://dx.doi.
org/10.1007/978-3-642-15877-3 8.
[60] Damljanovic D, Stankovic M, Laublet P. Linked Data-Based Concept
Recommendation: Comparison of Different Methods in Open Innovation
Scenario. In: Simperl E, Cimiano P, Polleres A, et al., editors. ESWC.
vol. 7295 of Lecture Notes in Computer Science. Springer; 2012. p. 24–
38. Available from: http://dblp.uni-trier.de/db/conf/esws/eswc2012.html#
DamljanovicSL12.
[61] Deldjoo Y, Elahi M, Cremonesi P, et al. Content-Based Video Recommenda-
tion System Based on Stylistic Visual Features. Journal on Data Semantics.
2016;p. 1–15.
[62] Elahi M, Deldjoo Y, Bakhshandegan Moghaddam F, et al. Exploring the Se-
mantic Gap for Movie Recommendations. In: Proceedings of the Eleventh
ACM Conference on Recommender Systems. ACM; 2017. p. 326–330.
[63] Deldjoo Y, Quadrana M, Elahi M, et al. Using Mise-En-Sc\ene Visual Fea-
tures based on MPEG-7 and Deep Learning for Movie Recommendation.
arXiv preprint arXiv:170406109. 2017;.
[64] Deldjoo Y, Elahi M, Quadrana M, et al. Using visual features based on
mpeg-7 and deep learning for movie recommendation. International Journal
of Multimedia Information Retrieval. 2018;.
[65] Lin K, Yang HF, Liu KH, et al. Rapid clothing retrieval via deep learning
of binary codes and hierarchical search. In: Proceedings of the 5th ACM on
International Conference on Multimedia Retrieval. ACM; 2015. p. 499–502.
[66] Bracher C, Heinz S, Vollgraf R. Fashion DNA: Merging content and
sales data for recommendation and article mapping. arXiv preprint
arXiv:160902489. 2016;.
“bigdata˙rs˙book2018”
2018/11/16
page 19
REFERENCES 19
[67] McAuley J, Targett C, Shi Q, et al. Image-based recommendations on styles
and substitutes. In: Proceedings of the 38th International ACM SIGIR
Conference on Research and Development in Information Retrieval. ACM;
2015. p. 43–52.
[68] Messina P, Dominquez V, Parra D, et al. Exploring Content-based Art-
work Recommendation with Metadata and Visual Features. arXiv preprint
arXiv:170605786. 2017;.
[69] Deldjoo Y, Elahi M, Cremonesi P, et al. Recommending movies based on
mise-en-scene design. In: Proceedings of the 2016 CHI Conference Ex-
tended Abstracts on Human Factors in Computing Systems. ACM; 2016. p.
1540–1547.
[70] Deldjoo Y, Elahi M, Quadrana M, et al. Toward Effective Movie Recom-
mendations Based on Mise-en-Sc`
ene Film Styles. In: Proceedings of the
11th Biannual Conference on Italian SIGCHI Chapter. ACM; 2015. p. 162–
165.
[71] Lehinevych T, Kokkinis-Ntrenis N, Siantikos G, et al. Discovering similari-
ties for content-based recommendation and browsing in multimedia collec-
tions. In: Signal-Image Technology and Internet-Based Systems (SITIS),
2014 Tenth International Conference on. IEEE; 2014. p. 237–243.
[72] Yang B, Mei T, Hua XS, et al. Online video recommendation based on
multimodal fusion and relevance feedback. In: Proceedings of the 6th ACM
international conference on Image and video retrieval. ACM; 2007. p. 73–
80.
[73] Zhao X, Li G, Wang M, et al. Integrating rich information for video rec-
ommendation with multi-task rank aggregation. In: Proceedings of the 19th
ACM international conference on Multimedia. ACM; 2011. p. 1521–1524.
[74] Canini L, Benini S, Leonardi R. Affective recommendation of movies based
on selected connotative features. Circuits and Systems for Video Technol-
ogy, IEEE Transactions on. 2013;23(4):636–647.
[75] Braunhofer M, Elahi M, Ricci F. User Personality and the New User Prob-
lem in a Context-Aware Point of Interest Recommender System. In: Infor-
mation and Communication Technologies in Tourism 2015. Springer; 2015.
p. 537–549.
[76] Nasery M, Elahi M, Cremonesi P. Polimovie: a feature-based dataset for
recommender systems. In: ACM RecSys Workshop on Crowdsourcing and
Human Computation for Recommender Systems (CrawdRec). vol. 3; 2015.
p. 25–30.
[77] Burger JM. Personality. Belmont, CA., USA: Wadsworth Publishing; 2010.
[78] John OP, Srivastava S. The Big Five trait taxonomy: History, measurement,
and theoretical perspectives.; 1999. Vol. 2, pp. 102 to 138. In: Handbook of
personality: Theory and research.
[79] Costa PT, McCrae RR. Revised NEO Personality Inventory (NEO PI-R)
and NEO Five-Factor Inventory (NEO FFI): Professional Manual. Psycho-
logical Assessment Resources; 1992.
“bigdata˙rs˙book2018”
2018/11/16
page 20
20 Big Data Recommender Systems: Recent Trends and Advances
[80] Hu R, Pu P. Enhancing collaborative filtering systems with personality in-
formation. In: Proceedings of the fifth ACM conference on Recommender
systems. RecSys ’11. New York, NY, USA: ACM; 2011. p. 197–204. Avail-
able from: http://doi.acm.org/10.1145/2043932.2043969.
[81] Nunes MASN, Hu R. Personality-based Recommender Systems: An
Overview. In: Proceedings of the 6th ACM Conference on Recommender
Systems; 2012. p. 5–6.
[82] Fern´
andez-Tob´
ıas I, Braunhofer M, Elahi M, et al. Alleviating the new
user problem in collaborative filtering by exploiting personality informa-
tion. User Modeling and User-Adapted Interaction. 2016;26(2-3):221–255.
[83] Schedl M, Zamani H, Chen CW, et al. Current challenges and visions in
music recommender systems research. International Journal of Multimedia
Information Retrieval. 2018;7(2):95–116.
[84] Rentfrow PJ, Gosling SD, et al. The do re mi’s of everyday life: The struc-
ture and personality correlates of music preferences. Journal of personality
and social psychology. 2003;84(6):1236–1256.
[85] Kosinski M, Stillwell D, Kohli P, et al. Personality and Website Choice. In:
Proceedings of the 3rd Annual ACM Web Science Conference. ACM; 2012.
.
[86] Odic A, Tkalcic M, Tasic JF, et al. Personality and social context: Impact
on emotion induction from movies. In: UMAP’13 Workshops; 2013. .
[87] Braunhofer M, Elahi M, Ge M, et al. Context Dependent Preference Ac-
quisition with Personality-Based Active Learning in Mobile Recommender
Systems. In: International Conference, HCI International 2014 (HCII’14).
Springer; 2014. .
[88] Tkalcic M, Kunaver M, Koˇ
sir A, et al. Addressing the new user problem
with a personality based user similarity measure. In: Proceedings of the 1st
International Workshop on Decision Making and Recommendation Accep-
tance Issues in Recommender Systems; 2011. p. 106.
[89] Nunes MASN. Recommender Systems based on Personality Traits: Could
Human Psychological Aspects Influence the Computer Decision-making
Process? VDM Verlag; 2009.
[90] Hu R, Pu P. A comparative user study on rating vs. personality quiz
based preference elicitation methods. In: Proceedings of the 14th inter-
national conference on Intelligent user interfaces. IUI ’09. New York, NY,
USA: ACM; 2009. p. 367–372. Available from: http://doi.acm.org/10.1145/
1502650.1502702.
[91] Roshchina A. TWIN Personality-based Recommender System. Institute of
Technology Tallaght, Dublin; 2012.
[92] Fern´
andez-Tob´
ıas I, Cantador I, Kaminskas M, et al. Cross-domain recom-
mender systems: A survey of the state of the art. In: Proceedings of the 2nd
Spanish Conference on Information Retrieval; 2012. p. 187–198.
[93] Winoto P, Tang TY. If you like the Devil Wears Prada the book, will you also
enjoy the Devil Wears Prada the movie? A study of cross-domain recom-
“bigdata˙rs˙book2018”
2018/11/16
page 21
REFERENCES 21
mendations. New Generation Computing. 2008;26(3):209–225. Available
from: http://dx.doi.org/10.1007/s00354-008-0041-0.
[94] Pagano R, Quadrana M, Elahi M, et al. Toward Active Learning in Cross-
domain Recommender Systems. arXiv preprint arXiv:170102021. 2017;.
[95] Abel F, Herder E, Houben GJ, et al. Cross-system user modeling and per-
sonalization on the Social Web. User Modeling and User-Adapted Inter-
action. 2013;23(2-3):169–209. Available from: http://dx.doi.org/10.1007/
s11257-012-9131-2.
[96] Berkovsky S, Kuflik T, Ricci F. Mediation of user models for enhanced per-
sonalization in recommender systems. User Modeling and User-Adapted In-
teraction. 2008;18(3):245–286. Available from: http://dx.doi.org/10.1007/
s11257-007-9042-9.
[97] Shapira B, Rokach L, Freilikhman S. Facebook single and cross domain
data for recommendation systems. User Modeling and User-Adapted Inter-
action. 2013;23(2-3):211–247. Available from: http://dx.doi.org/10.1007/
s11257-012-9128-x.
[98] Cantador I, Tob´
ıas IF, Berkovsky S, et al. Cross-domain recommender sys-
tems. In: Recommender Systems Handbook (2nd edition). Springer; 2015.
p. 919–959.
[99] Cremonesi P, Tripodi A, Turrin R. Cross-domain recommender systems. In:
Proceedings of the 11th International Conference on Data Mining Work-
shops; 2011. p. 496–503.
[100] Tiroshi A, Berkovsky S, Kˆ
aafar MA, et al. Cross social networks interests
predictions based on graph features. In: Proceedings of the 7th ACM Con-
ference on Recommender Systems; 2013. p. 319–322.
[101] Gao S, Luo H, Chen D, et al. Cross-domain recommendation via cluster-
level latent factor model. In: Proceedings of the 2013 European Confer-
ence on Machine Learning and Knowledge Discovery in Databases; 2013.
p. 161–176.
[102] Li B, Yang Q, Xue X. Can movies and books collaborate? cross-domain
collaborative filtering for sparsity reduction. In: Proceedings of the 21st In-
ternational Joint Conference on Artificial Intelligence; 2009. p. 2052–2057.
[103] Berkovsky S, Kuflik T, Ricci F. Distributed collaborative filtering with do-
main specialization. In: Proceedings of the 2007 ACM conference on Rec-
ommender systems. ACM; 2007. p. 33–40.
[104] Enrich M, Braunhofer M, Ricci F. Cold-Start management with cross-
domain collaborative filtering and tags. In: Proceedings of the 14th Interna-
tional Conference E-Commerce and Web Technologies; 2013. p. 101–112.
[105] Kohrs A, Merialdo B. Improving Collaborative Filtering For New-Users By
Smart Object Selection; 2001.
[106] Liu NN, Meng X, Liu C, et al. Wisdom of the better few: cold start recom-
mendation via representative based rating elicitation. In: Proceedings of the
fifth ACM conference on Recommender systems. ACM; 2011. p. 37–44.
[107] Teixeira IR, Carvalho FdATd, Ramalho G, et al. ActiveCP: A Method
for Speeding up User Preferences Acquisition in Collaborative Filtering
“bigdata˙rs˙book2018”
2018/11/16
page 22
22 Big Data Recommender Systems: Recent Trends and Advances
Systems. In: Proceedings of the 16th Brazilian Symposium on Artificial
Intelligence: Advances in Artificial Intelligence. SBIA ’02. London, UK,
UK: Springer-Verlag; 2002. p. 237–247. Available from: http://dl.acm.org/
citation.cfm?id=645853.669613.
[108] Rashid AM, Albert I, Cosley D, et al. Getting to know you: learning new
user preferences in recommender systems. In: Proceedings of the 7th inter-
national conference on Intelligent user interfaces. ACM; 2002. p. 127–134.
[109] Elahi M, Repsys V, Ricci F. Rating Elicitation Strategies for Collaborative
Filtering. In: Huemer C, Setzer T, editors. EC-Web. vol. 85 of Lecture Notes
in Business Information Processing. Springer; 2011. p. 160–171.
[110] Elahi M, Braunhofer M, Ricci F, et al. Personality-based active learning for
collaborative filtering recommender systems. In: Proceedings of the 13th In-
ternational Conference of the Italian Association for Artificial Intelligence.
Springer; 2013. p. 360–371.
[111] Zhou K, Yang SH, Zha H. Functional matrix factorizations for cold-start
recommendation. In: Proceedings of the 34th international ACM SIGIR
conference on Research and development in Information Retrieval. SIGIR
’11. New York, NY, USA: ACM; 2011. p. 315–324. Available from: http:
//doi.acm.org/10.1145/2009916.2009961.
[112] Kaminskas M, Fernndez-Tobas I, Ricci F, et al. Knowledge-based music
retrieval for places of interest. In: Liem CCS, Mller M, Tjoa SK, et al.,
editors. MIRUM; 2012. p. 19–24. Available from: http://dblp.uni-trier.de/
db/conf/mm/mirum2012.html#KaminskasFRC12.
[113] Passant A. dbrec - Music Recommendations Using DBpedia. In: Pro-
ceedings of the 9th International Semantic Web Conference (ISWC 2010).
Springer; 2010. p. 209–224.
[114] Ristoski P, Menca EL, Paulheim H. A Hybrid Multi-strategy Recommender
System Using Linked Open Data. In: Presutti V, Stankovic M, Cambria
E, et al., editors. SemWebEval@ESWC. vol. 475 of Communications in
Computer and Information Science. Springer; 2014. p. 150–156. Avail-
able from: http://dblp.uni-trier.de/db/conf/esws/semwebeval2014.html#
RistoskiMP14.
[115] Ostuni VC, Noia TD, Sciascio ED, et al. Top-N recommendations from
implicit feedback leveraging linked open data. In: Yang Q, King I, Li Q,
et al., editors. RecSys. ACM; 2013. p. 85–92. Available from: http://dblp.
uni-trier.de/db/conf/recsys/recsys2013.html#OstuniNSM13.
[116] Zhang Y, Wu H, Sorathia VS, et al. Event Recommendation in So-
cial Networks with Linked Data Enablement. In: Hammoudi S, Ma-
ciaszek LA, Cordeiro J, et al., editors. ICEIS (2). SciTePress; 2013. p.
371–379. Available from: http://dblp.uni-trier.de/db/conf/iceis/iceis2013-2.
html#ZhangWSP13.
[117] Deldjoo Y, Elahi M, Cremonesi P, et al. How to Combine Visual Fea-
tures with Tags to Improve Movie Recommendation Accuracy? In: In-
ternational Conference on Electronic Commerce and Web Technologies.
Springer; 2016. p. 34–45.
“bigdata˙rs˙book2018”
2018/11/16
page 23
REFERENCES 23
[118] Cremonesi P, Elahi M, Garzotto F. Interaction design patterns in recom-
mender systems. In: Proceedings of the 11th Biannual Conference on Italian
SIGCHI Chapter. ACM; 2015. p. 66–73.
