Cross-Domain Recommendations with Overlapping Items

Denis Kotkov, Shuaiqiang Wang and Jari Veijalainen

University of Jyvaskyla, Dept. of Computer Science and Information Systems,

P.O.Box 35, FI-40014 Jyvaskyla, Finland

Keywords:

Recommender Systems, Cross-Domain Recommendations, Collaborative Filtering, Content-Based Filtering,

Data Collection.

Abstract:

In recent years, there has been an increasing interest in cross-domain recommender systems. However, most

existing works focus on the situation when only users or users and items overlap in different domains. In

this paper, we investigate whether the source domain can boost the recommendation performance in the target

domain when only items overlap. Due to the lack of publicly available datasets, we collect a dataset from

two domains related to music, involving both the users’ rating scores and the description of the items. We

then conduct experiments using collaborative ﬁltering and content-based ﬁltering approaches for validation

purpose. According to our experimental results, the source domain can improve the recommendation perfor-

mance in the target domain when only items overlap. However, the improvement decreases with the growth

of non-overlapping items in different domains.

1 INTRODUCTION

Recommender systems use past user behavior to sug-

gest items interesting to users (Ricci et al., 2011). An

item is a piece of information that refers to a tangible

or digital object, such as a good, a service or a process

that a recommender system suggests to the user in an

interaction through the Web, email or text message.

The majority of recommender systems suggest

items based on a single domain. In this paper, the

term domain refers to “a set of items that share certain

characteristics that are exploited by a particular rec-

ommender system” (Fern

andez-Tob

ıas et al., 2012).

These characteristics are items’ attributes and users’

ratings.

However, the single domain recommender sys-

tems often suffer from data sparsity and cold start

problems. In order to overcome these limitations it

is possible to consider data from different domains.

Recommender systems that take advantage of mul-

tiple domains are called cross-domain recommender

systems (Fern

andez-Tob

ıas et al., 2012; Cantador and

Cremonesi, 2014).

In this paper, we consider a cross-domain recom-

mendation task (Cantador and Cremonesi, 2014), that

requires one target domain and at least one source do-

main. The former refers to the domain that suggested

items are picked from, and similarly the latter refers

to the additional domain that contains auxiliary infor-

mation.

Cross-domain recommender systems can be clas-

siﬁed based on domain levels (Cantador and Cre-

monesi, 2014):

• attribute level - items can be assigned to different

domains according to their descriptions. One may

contain jazz music, while another may consist of

pop audio recordings;

• type level - items may have different types, but

share common attributes. Movie and book do-

mains have common genres, such as drama, com-

edy and horror, while movies and books have dif-

ferent types;

• item level - items from different domains may

have completely different attributes and types.

Songs and books might not share any common at-

tributes;

• system level - items may belong to different rec-

ommender systems, have the same type and share

many common attributes. For example, movies

from IMDb

and MovieLens

may belong to dif-

ferent domains.

Depending on whether overlapping occurs in the

set of users or items (Cremonesi et al., 2011), there

http://www.imdb.com/

https://movielens.org/

Kotkov, D., Wang, S. and Veijalainen, J.

Cross-Domain Recommendations with Overlapping Items.

In Proceedings of the 12th International Conference on Web Information Systems and Technologies (WEBIST 2016) - Volume 2, pages 131-138

ISBN: 978-989-758-186-1

131

are four situations that enable cross-domain recom-

mendations: a) no overlap between items and users,

b) user sets of different domains overlap, c) item sets

overlap, and d) item and user sets overlap.

In this work, we investigate whether the source

domain improves the recommendation performance

in the target domain on system level in the situation

when only items overlap. The idea behind the paper

is as follows. Traditional cross-domain recommender

systems utilize overlapping users to discover addi-

tional interests of users, leading to the improvement

of the recommendation diversity. When the items

overlap, the source domain lets detect more accurate

similarities between items, which should positively

result in recommendation performance in the target

domain.

Due to the lack of publicly available datasets for

cross-domain recommender systems with overlapping

items (Berkovsky et al., 2008; Kille et al., 2013), we

collected data from Vkontakte

(VK) – Russian on-

line social network (OSN) and Last.fm

(FM) – mu-

sic recommender service. We then matched VK and

FM audio recordings and developed the cross-domain

recommender system that suggests VK recordings to

VK users based on data from both domains. Each au-

dio recording is represented by its metadata excluding

the actual audio ﬁle. VK recordings thus represent

the target domain, while the source domain consists

of FM recordings. VK and FM recordings share titles

and artists, but have different user ratings and other

attributes.

In order to address the research question and illus-

trate the potential of additional data, we chose sim-

ple but popular recommendation algorithms to con-

duct experiments for validation: collaborative ﬁlter-

ing based on users’ ratings and content-based ﬁltering

based on the descriptions of the items.

Our results indicate that the source domain can

improve the recommendation performance in the tar-

get domain. Furthermore, with the growth of non-

overlapping items in different domains, the improve-

ment of recommendation performance decreases.

This paper thus has the following contributions:

• we initially investigate the cross-domain recom-

mendation problem in the situation when only

items overlap;

• we collect a novel dataset to conduct the experi-

ments for addressing the research question.

The paper might be useful in real life scenarios.

For example, according to our results, the perfor-

mance of a recommender system lacking user rat-

http://vk.com/

http://last.fm/

ings to achieve an acceptable performance can be im-

proved using ratings collected from another recom-

mender system that suggests items of the same type.

However, the performance might decrease if the rec-

ommender systems do not have enough overlapping

items.

The rest of the paper is organized as follows. Sec-

tion 2 overviews related works. Section 3 describes

the datasets used to conduct experiments. Section 4 is

dedicated to recommendation approaches, while sec-

tion 5 describes conducted experiments. Finally, sec-

tion 6 draws ﬁnal concussions.

2 RELATED WORKS

Most existing approaches consider additional infor-

mation about users to boost the recommendation per-

formance. For example, one of the ﬁrst studies dedi-

cated to cross-domain recommender systems investi-

gated the effectiveness of source domains with over-

lapping users (Winoto and Tang, 2008). In the exper-

iment, undergraduates from a local university were

asked to rate items from different domains, such as

movies, songs and books. The authors measured rec-

ommendation performance in different domain com-

binations and concluded that source domains decrease

the recommendation performance, but may improve

the diversity of recommendations.

In contrast, other studies demonstrated that source

domains can boost the recommendation performance

in the target domain in situations when users or both

users and items overlap. For example, Sang demon-

strated the feasibility of utilizing the source domain.

The study was conducted on a dataset collected from

Twitter

and YouTube

. The author established rela-

tionships between items from different domains using

topics (Sang, 2014). Similarly to Sang, Shapira et al.

also linked items from different domains, where 95

participants rated movies and allowed the researches

to collect data from their Facebook pages. The re-

sults suggested that additional domains improve the

recommendation performance (Shapira et al., 2013).

Another study with positive results was conducted by

Abel et al. The dataset contained information related

to the same users from 7 different OSNs (Abel et al.,

2013). Sahebi et al. demonstrated the usefulness

of recommendations based on additional domains to

overcome cold start problem (Sahebi and Brusilovsky,

2013).

Most works on cross-domain recommender sys-

tems focus on the situation when users or both users

https://twitter.com/

https://www.youtube.com/

WEBIST 2016 - 12th International Conference on Web Information Systems and Technologies

132

Figure 1: Data collection chart.

and items of several domains overlap (Cantador and

Cremonesi, 2014). However, to the best of our knowl-

edge, the efforts on the impact of source domains on

the target domain with only overlapping items involv-

ing a real cross-domain dataset is very limited.

3 DATASETS

Due to the lack of publicly available datasets for

cross-domain recommender systems with overlapping

items (Berkovsky et al., 2008; Kille et al., 2013) we

collected data from VK and FM. The construction of

the dataset included three phases (Figure 1): 1) VK

recordings collection, 2) duplicates matching, and 3)

FM recordings collection.

3.1 VK Recordings Collection

The VK interface provides the functionality to add

favorite recordings to users’ pages. By generating

random user ids we collected disclosed VK users’

favorite audio recordings using VK API. Our VK

dataset consists of 97, 737 (76, 177 unique) audio

recordings added by 864 users.

Each VK user is allowed to share any audio or

video recording. The interface of the OSN provides

the functionality to add favorite recordings to the

users page. VK users are allowed not only to add

favorite audio recordings to their pages, but also to

rename them. The dataset thus contains a noticeable

number of duplicates with different names. To assess

this number we randomly selected 100 VK recordings

and manually split them into three categories:

• correct names - the name of the recording is cor-

rectly written without any grammatical mistakes

or redundant symbols;

• misspelled names - the name is guessable, even

if the name of the recording is replaced with

the combination of artist and recording name or

lyrics;

• meaningless names – the name does not contain

any information about the recording. For exam-

ple, “unknown” artist and “The song” recording.

Out of 100 randomly selected recordings we detected

14 misspelled and 2 meaningless names. The example

can be seen from table 1.

Table 1: Examples of recordings.

Artist name Recording name

Correct names

Beyonce Halo

Madonna Frozen

Misspelled

Alice DJ Alice DJ - Better of Alone.mp3

Reamonn Oh, tonight you kill me with your smile

l Lady Gaga Christmas Tree

Meaningless

Unknown classic

Unknown party

3.2 Duplicates Matching

In order to match misspelled recordings, we de-

veloped a duplicate matching algorithm that detects

duplicates based on recordings’ names, mp3 links

and durations. The algorithm compares recordings’

names based on Levenshtein distance and the number

of common words excluding stop words.

We then removed some popular meaningless

recordings such as “Unknown”, “1” or “01”, because

they represent different recordings and do not indicate

users’ preferences. Furthermore, some users assign

wrong popular artists’ names to the recordings. To re-

strict the growth of this kind of mistakes, the matching

algorithm considers artists of the duplicate recordings

to be different. By using the presented matching ap-

proach, the number of unique recordings decreased

from 76, 177 to 68,699.

3.3 FM Recordings Collection

In order to utilize the source domain we collected

FM recordings that correspond to 48, 917 selected

VK recordings that were added by at least two users

or users that have testing data. Each FM record-

ing contains descriptions such as FM tags added by

FM users. FM tags indicate additional information

such as genre, language or mood. Overall, we col-

lected 10, 962 overlapping FM recordings and 20, 214

(2, 783 unique) FM tags.

It is also possible to obtain FM users who like a

certain recording (top fans). For each FM recording,

we collected FM users who like at least one more FM

recording from our dataset according to the distribu-

tion of VK users among those recordings. In fact,

some unpopular FM recordings are missing top fans.

We thus collected 17, 062 FM users, where 7, 083 of

them like at least two recordings from our database.

In this work, we constructed three datasets. Each

of them includes the collected FM data and different

parts of the VK data:

Cross-Domain Recommendations with Overlapping Items

133

• 0% - the dataset contains only overlapping record-

ings rated by VK and FM users;

• 50% - the dataset contains overlapping recordings

and the half of randomly selected VK recordings

that do not correspond to FM recordings;

• 100% - the dataset contains all collected VK and

FM recordings.

The statistics of the datasets are presented in ta-

ble 2. The number of VK users varies in different

dataset, due to the lack of ratings after removing non-

overlaping VK recordings.

4 RECOMMENDATION

APPROACHES

In order to emphasize the importance of additional

data we implemented simple, but popular collabora-

tive ﬁltering and content-based ﬁltering algorithms.

4.1 Item-based Collaborative Filtering

Each recording is represented as a vector in the mul-

tidimensional feature space, where each feature is a

user’s choice. VK recording is represented as follows:

= (u

1, j

, u

2, j

, ..., u

n, j

), where u

k, j

equals to 1 if VK

user k picks VK recording j and 0 otherwise. The

representation changes if we consider the FM users:

vk, f m

= (u

1, j

, u

2, j

, ..., u

n, j

, u

f m

1, j

, u

f m

2, j

, ..., u

f m

n, j

In order to rank items in the suggested list we

use sum of similarities of recordings (Ekstrand et al.,

2011):

score(u

, i

) =

∑

∈I

sim(i

, i

), (1)

where I

is the set of items picked by u

user. We

use conditional probability as similarity measure (Ek-

strand et al., 2011):

p(i

, i

) =

Freq(i

∧ i

)

Freq(i

) · Freq(i

)

, (2)

where Freq(i

) is the number of users that liked item

, while Freq(i

∧i

) is the number of users that liked

both items i

and i

. The parameter α is a demping

factor to decrese the similarity for popular items. In

our experiments α = 1.

It is worth mentioning that item vectors based on

FM users contain remarkably more dimensions than

vectors based on VK users. In order to alleviate the

problem we compared recordings using the following

rule:

sim(i

, i

) =











p(i

, i

), ∃i

∧ ∃i

∧

(@i

f m

∨ @i

f m

)

p(i

f m

, i

f m

), ∃i

f m

∧ ∃i

f m

∧

(@i

∨ @i

)

p(i

vk, f m

, i

vk, f m

), ∃i

∧ ∃i

∧

∃i

f m

∧ ∃i

f m

. (3)

We compare items in each pair using only do-

mains that contain users’ ratings for both items.

4.2 Content-based Filtering

In a content-based approach similarly to an item-

based approach each recording is represented as a

vector, but each dimension corresponds to an attribute

of the item. In our case, these attributes are VK FM

artists and FM tags. It is worth mentioning that FM

and VK artists correspond to each other.

An audio recording thus is represented as follows:

= (a

1, j

, a

2, j

, ..., a

d, j

), where a

k, j

equals to 1 if the

recording i

is performed by the artist a

and 0 oth-

erwise. The user then can be represented similarly:

= (a

1, j

, a

2, j

, ..., a

d, j

), where a

k, j

equals to 1 if user

k picks the recording perfromed by the artist a

and 0

otherwise.

The representation chages if we consider FM tags:

= (w

1, j

, w

2, j

, ..., w

q, j

), where w

k, j

corresponds to

the term frequencyinverse document frequency (Lops

et al., 2011). The user vector then is denoted as fol-

lows: u

= (t

1, j

2, j

, ..., t

q, j

), where t

k, j

is a number of

recodings that have tag t

and are picked by user u

The recommender system compares audio record-

ings’ vectors and a user vector using cosine similar-

ity (Ekstrand et al., 2011). First, the suggested list

is sorted according to the similarity based on artists.

Second, list fragments that consist of items with the

same artists’ similarity are sorted according to the FM

tag similarity.

5 EXPERIMENTS

In this section, we conduct experiments to demon-

strate whether the source domain improves the rec-

ommendation performance in the target domain when

only items overlap.

5.1 Evaluation Metrics

We used precision@K, recall@K, mean average pre-

cision (MAP) and normalized discounted cumula-

tive gain (NDCG) to evaluate our approaches (Zhao,

WEBIST 2016 - 12th International Conference on Web Information Systems and Technologies

134

Table 2: The Statistics of the Datasets.

0% 50% 100%

VK FM VK FM VK FM

Users 661 7,083 850 7,083 864 7,083

Ratings 14,207 40,782 62,435 40,782 96,737 40,782

Items 4,605 4,605 39,831 4,605 68,699 4,605

Artists 1,986 1,986 19,930 1,986 31,861 1,986

Tags - 20,167 - 20,167 - 20,167

2013), as these metrics are the most popular in infor-

mation retrieval. Precision@K, recall@K and mean

average precision (MAP) are used to assess quality of

recommended lists with binary relevance. Binary rel-

evance requires each item to be relevant or irrelevant

for a particular user. As in our case a user can only

indicate relevance of a recording by adding it to her

page, we regarded added recordings as equally rele-

vant for a user. We regarded the rest recordings as

irrelevant. Precision@K and recall@K for a speciﬁc

user are calculated as follows:

@K =

(K)

, (4)

@K =

(K)

, (5)

where r

(K) is the number of items relevant for user

u in the ﬁrst K results, while r

indicates the number

of relevant items in the list recommended to user u.

Overall precision@K and recall@K are average val-

ues.

Precision@K =

||U||

∑

u∈U

@K, (6)

Recall@K =

||U||

∑

u∈U

@K, (7)

where U is a set of evaluated users. MAP then can be

calculated in the following way:

MAP =

||U||

∑

u∈U

∑

i=1

u,i

· P

, (8)

where r

u,i

= 1 if an item at position i in the recom-

mended list is relevant for user u and r

u,i

= 0 other-

wise. In our experiments h was set to 30.

We also evaluated our approaches using NDCG

arvelin and Kek

ainen, 2002). The metric consid-

ers positions of items in recommended lists and multi-

ple levels of relevance. We employed NDCG to mea-

sure the quality of recommendations with binary rel-

evance. The metric is calculated as follows:

NDCG

@K = Z

∑

i=1

(

u,i

− 1, i = 1

u,i

−1

log

(i)

, i > 1

, (9)

NDCG@K =

||U||

∑

u∈U

NDCG

@K, (10)

where Z

is the normalization constant.

5.2 Results

Following the datasets sampling strategy in (Ekstrand

et al., 2011), we split each of our datasets into training

and test datasets. In particular, we selected 40% of

the users who rated the most VK recordings, and then

chose 30% of their ratings as the testing dataset. We

then regarded the rest ratings as the training dataset.

We used ofﬂine evaluation to compare results of

proposed methods with baselines. The recommender

system suggested 30 popular VK recordings to each

testing VK user excluding recordings that the user has

already added in the training set. In each approach the

recommendation list consists of the same items. We

chose popular items for evaluation, due to the high

probability that users have seen them already.

In this study, we demonstrate the performance

improvement resulting from the source domain with

three simple but popular algorithms: (1) POP, (2) Col-

laborative Filtering (CF), and (3) Content-based Fil-

tering (CBF). In particular, POP is a non-personalized

recommendation algorithm, which orders items in the

suggested list according to their popularity in the VK

dataset. For the CF and the CBF algorithms, we ob-

tained two performance results based on only VK and

VK+FM data.

• POP - ordering items according to their popular-

ity using the VK dataset.

• CF(VK) - item-based collaborative ﬁltering using

the VK dataset.

• CF(VK+FM) - item-based collaborative ﬁltering

using VK and FM datasets.

• CBF(VK) - content-based ﬁltering using the VK

dataset.

• CBF(VK+FM) - content-based ﬁltering using

VK and FM datasets.

Figures 2, 3 and 4 demonstrate the experimental

results based on three datasets presented in Section 3.

From the ﬁgures we can observe that:

1. The source domain can improve the recommenda-

tion performance in the target domain when only

items overlap. For the 0% dataset, the CF algo-

rithm achieves 0.0216, 0.0273, 0.0139 and 0.0287

Cross-Domain Recommendations with Overlapping Items

135

0,01

0,015

0,02

0,025

0,03

0,035

POP

CF(VK)

CF(VK+FM)

CBF(VK)

CBF(VK+FM)

(a) Precision@K (0%)

0,01

0,02

0,03

0,04

0,05

0,06

POP

CF(VK)

CF(VK+FM)

CBF(VK)

CBF(VK+FM)

(b) Recall@K (0%)

0,015

0,017

0,019

0,021

0,023

0,025

0,027

POP

CF(VK)

CF(VK+FM)

CBF(VK)

CBF(VK+FM)

0,002

0,004

0,006

0,008

0,01

0,012

0,014

POP

CF(VK)

CF(VK+FM)

CBF(VK)

CBF(VK+FM)

(d) Recall@K (50%)

0,015

0,02

0,025

0,03

0,035

POP

CF(VK)

CF(VK+FM)

CBF(VK)

CBF(VK+FM)

(e) Precision@K (100%)

0,002

0,003

0,004

0,005

0,006

0,007

0,008

0,009

POP

CF(VK)

CF(VK+FM)

CBF(VK)

CBF(VK+FM)

(f) Recall@K (100%)

Figure 2: Precision@K and Recall@K for experiments conducted using datasets with different fractions of non-overlapping

items.

in terms of precision@10, recal@10, MAP and

NDCG@10 based on VK dataset, while these

numbers are 0.0297, 0.0372, 0.0179 and 0.0387

based on VK+FM dataset, making the improve-

ment of 37.5%, 36.3%, 28.8% and 34.8%, respec-

tively. Similar improvements can be observed for

the CBF algorithm.

2. The improvement declines with the growth of

non-overlapping items in different domains. For

example, the improvement of CBF in terms of

NDCG@10 decreases as follows: 20.1%, 5.4%

and 5.0% using 0%, 50% and 100% datasets,

respectively. For the CF algorithm, the declin-

ing trend is even sharper. The source domain

decreases the performance of the CF algorithm

by 11.8% and 7.0% in terms of NDCG@5 and

NDCG@10 respectively using 100% dataset. A

similar trend can be observed for other numbers

of ﬁrst K results and evaluation metrics.

3. CF(VK) and CBF(VK) perform worse than POP

in different cases, especially using the dataset

that contains only overlapping recordings (0%).

CF(VK) algorithm outperforms the popularity

baseline with the increase of non-overlapping

recordings. CF(VK) achieves 0.0139, while POP

outperforms them with the number 0.0180 in

terms of MAP using 0% dataset. For 100%

dataset the situation is opposite. POP achieves

0.0029, while CF(VK) reaches 0.0031. POP out-

performs CBF(VK) algorithm in most cases. For

0% and 100% datasets, CBF(VK) performs 1.9%

and 8.4% worse than POP in terms of MAP, re-

spectively.

Observation 1 illustrates the global correlation of

WEBIST 2016 - 12th International Conference on Web Information Systems and Technologies

136

0,01

0,012

0,014

0,016

0,018

0,02

POP

CF(VK)

CF(VK+FM)

CBF(VK)

CBF(VK+FM)

(a) MAP (0%)

0,003

0,0032

0,0034

0,0036

0,0038

0,004

0,0042

0,0044

0,0046

POP

CF(VK)

CF(VK+FM)

CBF(VK)

CBF(VK+FM)

(b) MAP (50%)

0,002

0,0022

0,0024

0,0026

0,0028

0,003

0,0032

0,0034

POP

CF(VK)

CF(VK+FM)

CBF(VK)

CBF(VK+FM)

Figure 3: MAP (30 recommendations) for experiments conducted using datasets with different fractions of non-overlapping

items.

0,015

0,02

0,025

0,03

0,035

0,04

0,045

50%

100%

POP

CF(VK)

CF(VK+FM)

CBF(VK)

CBF(VK+FM)

(a) NDCG@5

0,015

0,02

0,025

0,03

0,035

0,04

0,045

50%

100%

POP

CF(VK)

CF(VK+FM)

CBF(VK)

CBF(VK+FM)

(b) NDCG@10

Figure 4: NDCG@K for experiments conducted using datasets with different fractions of non-overlapping items.

users’ preferences in different domains (Winoto and

Tang, 2008; Fern

andez-Tob

ıas et al., 2012). Al-

though, the data belongs to different domains, users’

ratings from the source domain indicate similarities

between items that improve the recommendation per-

formance in the target domain.

Observation 2 supports the claim (Fern

andez-

Tob

ıas et al., 2012), that the improvement cased by

the source domain rises with the growth of the overlap

between target and source domains. The decrease in

the recommendation performance of the CF algorithm

with the FM data is caused by the different lengths of

item vectors in source and target domains, where vec-

tors of FM items contain signiﬁcantly more dimen-

sions than vectors of VK items.

In observation 3, the non-personalized algorithm

POP outperforms both the personalized algorithms

in different cases. CF algorithm performs worse

than POP due to data sparsity, which is alleviated by

adding more VK recordings to the dataset. Low per-

formance of CBF is caused by the poor quality of item

descriptions, as in the VK dataset items are described

with artists only.

Figures 2, 3 and 4 demonstrate four evaluation

metrics that are not always consistent. However, the

described observations can still be notices.

6 CONCLUSION

In this paper we investigated cross-domain recom-

mendations in the situation when only items overlap

on system level. We collected data from VK and FM

and built three datasets that contain different fractions

of non-overlapping items from source and target do-

mains. We then conducted experiments using collab-

orative ﬁltering and content-based ﬁltering algorithms

to demonstrate the importance of additional data.

According to our results, the source domain can

boost the recommendation performance in the target

domain when only items overlap resulting from the

correlation of users’ preferences among different do-

mains (Winoto and Tang, 2008). However, similarly

to (Fern

andez-Tob

ıas et al., 2012) our results indi-

cated that the more items overlap in source and target

domains with respect to the whole dataset the higher

the improvement.

ACKNOWLEDGEMENT

The research at the University of Jyv

askyl

a was per-

formed in the MineSocMed project, partially sup-

ported by the Academy of Finland, grant #268078.

Cross-Domain Recommendations with Overlapping Items

137

REFERENCES

Abel, F., Herder, E., Houben, G.-J., Henze, N., and Krause,

D. (2013). Cross-system user modeling and personal-

ization on the social web. User Modeling and User-

Adapted Interaction, 23(2-3):169–209.

Berkovsky, S., Kuﬂik, T., and Ricci, F. (2008). Mediation

of user models for enhanced personalization in recom-

mender systems. User Modeling and User-Adapted

Interaction, 18(3):245–286.

Cantador, I. and Cremonesi, P. (2014). Tutorial on cross-

domain recommender systems. In Proceedings of the

8th ACM Conference on Recommender Systems, Rec-

Sys ’14, pages 401–402, New York, NY, USA. ACM.

Cremonesi, P., Tripodi, A., and Turrin, R. (2011). Cross-

domain recommender systems. In Data Mining Work-

shops (ICDMW), 2011 IEEE 11th International Con-

ference on, pages 496–503.

Ekstrand, M. D., Riedl, J. T., and Konstan, J. A. (2011).

Collaborative ﬁltering recommender systems. Foun-

dations and Trends in Human-Computer Interaction,

4(2):81–173.

Fern

andez-Tob

ıas, I., Cantador, I., Kaminskas, M., and

Ricci, F. (2012). Cross-domain recommender sys-

tems: A survey of the state of the art. In Spanish

Conference on Information Retrieval.

arvelin, K. and Kek

ainen, J. (2002). Cumulated gain-

based evaluation of ir techniques. ACM Transactions

on Information Systems, 20(4):422–446.

Kille, B., Hopfgartner, F., Brodt, T., and Heintz, T. (2013).

The plista dataset. In Proceedings of the 2013 Inter-

national News Recommender Systems Workshop and

Challenge, NRS ’13, pages 16–23, New York, NY,

USA. ACM.

Lops, P., de Gemmis, M., and Semeraro, G. (2011).

Content-based recommender systems: State of the

art and trends. In Recommender Systems Handbook,

pages 73–105. Springer US.

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

to Recommender Systems Handbook. Springer US.

Sahebi, S. and Brusilovsky, P. (2013). Cross-domain collab-

orative recommendation in a cold-start context: The

impact of user proﬁle size on the quality of recommen-

dation. In User Modeling, Adaptation, and Personal-

ization, volume 7899 of Lecture Notes in Computer

Science, pages 289–295. Springer Berlin Heidelberg.

Sang, J. (2014). Cross-network social multimedia comput-

ing. In User-centric Social Multimedia Computing,

Springer Theses, pages 81–99. Springer Berlin Hei-

delberg.

Shapira, B., Rokach, L., and Freilikhman, S. (2013). Face-

book single and cross domain data for recommenda-

tion systems. User Modeling and User-Adapted Inter-

action, 23(2-3):211–247.

Winoto, P. and Tang, T. (2008). If you like the devil wears

prada the book, will you also enjoy the devil wears

prada the movie? a study of cross-domain recommen-

dations. New Generation Computing, 26(3):209–225.

Zhao, Y.-L. (2013). Community Learning in Location-

based Social Networks. Thesis. Ph.D.

WEBIST 2016 - 12th International Conference on Web Information Systems and Technologies

138