Improving recommendation of leisure activity using social factor in cross-domain approach Текст научной статьи по специальности «Экономика и бизнес»

ISSN 1992-6502 (Print)_TB&C/ftVH/U/IC ^¡jf/^Q^^lj _ISSN 2225-2789 (Online)

Vol. 18, no. 5 (66), pp. 91-95, 2014 http://journal.ugatu.ac.ru

UDC 004.65

Improving recommendation of leisure activity

using social factor in cross-domain approach m. r. Badretdinov 1, t. r. Badretdinov 2, g. a. Makeev 3, f. Casati 4

1 timars@list.ru, 2 maeglin777@mail.ru, 3 greg@ufanipi.ru, 4 casati@disi.unitn.it

1-3 Ufa State Aviation Technical University, Russia 4 University of Trento, Italy

Received 18.07.2014

Abstract. In this paper we want to analyze how social factor such as friendship on Facebook can influence cross-domain recommendation results. For this we analyzed preferences of people from several cities around the world about various types of leisure activities, taking into consideration different purposes for which activity is performed.

Keywords: recommender systems; collaborative filtering; cross-domain recommendation; social recommendation.

INTRODUCTION

Nowadays, we see tremendous amount of options when purchasing movies, books or looking for leisure activity. Despite the overwhelming number of options we are exposed to, we are still missing out a plenty of opportunities, but not because we don't want to, but because we are not aware of the possibility. This raises a need for intelligent systems providing personalized service with respect to users' needs and interests, represented by user models. Recommender systems, which become more popular and widespread, can be applied for solving this problem. Collaborative filtering is one of the most popular and widely used recommender systems approaches. In collaborative filtering recommendations are based on users' behaviour, i. e. users are similar if they have similar preferences, if they like the same options. Thus we make decisions using not the content of the available options, but in users' attitude to these options.

Different recommender systems were successfully applied in such well-known digital services as Amazon, Netflix, MovieLens, Last.fm, Pan-dora.com and many others. There are also a number of online services such as TripAdvisor, Foursquare, Yelp and Evenbrite providing different types of suggestions for activities to perform in leisure time, events or places to visit. Although such services help people to focus attention to a reduced number of events, in most cases people still have the feeling of missing out interesting activities [1]. In recent

researches [1, 2] it was shown that taking into consideration social factor such as friendship on Face-book improves performance of leisure activity recommendation in comparison with user based collaborative filtering approach using k-most-similar users. Also it was shown [2] that information about users' preferences in one leisure activity domain can be used to make prediction of users' preferences in another leisure activity domain even without any information of user's preferences in second leisure activity domain, that helps to solve the cold start problem of collaborative filtering and thus provide better recommendations, extending the knowledge-base to different leisure activity domains. In this study we want to find out how social factor influence performance of cross-domain recommendation in comparison with user based collaborative filtering approach using k-most-similar users.

FORMAL EXPERIMENT DEFINITION

Let U be the set of all users and P the set of all places of a possible activity. Let A be the set of all activities like drinking aperitivo in a bar, having dinner at a restaurant, drinking some beer in a pub or dancing in a club. Let G be the set of goals that can be accomplished with an activity. For example, goal can be something like achieving best price/quality ratio [2].

L ik e d (u, p,a,g )EUxPxAxG - be all the places user rated positively for a given activity and a given goal.

D is Ike d (u, p,a,g )eUxPxAxG - be all the places user rated negatively for a given activity and a given goal.

Rated(u,p, a, g) = Liked(u,p,a,g) U Disliked(u,p,a,g),

Known(u,a) = {pePl3geG,Rated(u,p,a,g)j,

Unknown(u,a) = P\Known(u, a).

In the studies we will consider two relations between users: similarity and friendship.

and are

Facebook friends.

Similarity is the ratio of similarly rated activities from co-rated set of activities to a number of all co-rated activities. In other words shows how much preferences of one user coincide with preferences of another user [2].

. ; \\Corated(u,u',a)\\

\\Known(u,a) n Unknown(u,a)\\'

Corated(u,u' ,a) = {peP\geG,

Liked(u,p,a, g) C\Liked{u',p, a,g) U

Disliked(u,p, a, g) n Disliked(u',p, a, g)}.

Recommendation of places p to a user u to perform an activity a with a goal g [2]:

1) Rec(u,a,g,k) £ Unknown(u,a).

2) | | R ec (u,a,g,k) \ \ = k.

3) Vp £ Rec(u,a,g,k)

Vp' £ (Unknown(u,a)\Rec(u,a, g.kf) (score(Net(u),p,a,g) > score(Net(u),p' , a,g)) ,

where scoring function for a place on a network to perform an activity with a goal is defined as average rating of users :

score(p, Net(u), a, g)

\\Likes(p,Net(u),a,g)\\ — \\Dislikes(p,Net(u),a, g)\\

= ||jVet(u)||

Likes(p, Net(u), a, g)

= {u' £ Net(u)\Liked(u',p,a,g)},

Dislikes (p, Net{u), a, g)

= {u' £ Net(u)\Disliked(u',p, a,g)}.

In this study we focused on understanding whether recommendation across different activities coming from similar friends gives better performance than recommendation coming from similar friends. Thus we defined the networks for similar users and similar friends

for recommendation across different

activities:

Netsu(u) = {u' e U\3a' £ A,sim(u,u',a') > 5},

Netsf(u) = [u' £ U\3a' £ A,FriendOf{u,u') n sim(u,u',a') > 5}.

INITIAL DATA

Three different cities around the world were considered: Trento (Italy), Asunción (Paraguay) and Tomsk (Russia). In each city ratings were acquired not only for restaurants but also for another activity that is usually done before or after going out for dinner: drinking aperitif in a bar in Trento (Italy), drinking some beer in a pub or bar in Asunción (Paraguay), dancing in a club in Tomsk (Russia). For each place people specified four different marks according to different goals: one mark was dedicated to the price / quality ratio and the other three were related to the different types of companions people can spend their leisure time with, which are tourists, friends and their partner [2]. Also information about friendship on Facebook was obtained. In this study we used data gathered in Trento University with help of service ComeAlong. Gathered data contain a total of 9820 ratings from 162 local people on 353 restaurants and 85 places for second activity (Table 1).

Table 1

Gathered data

Table 2

Co-rated activities number for Activity1 (Visiting restaurant)

Number of co-rated activities Number of users on the average

Trento (Italy) Asunción (Paraguay) Tomsk (Russia)

0 8 40 2

1 10 25 2

2 6 12 2

3 5 6 1

4 2 3 1

5 1 1 1

Trento (Italy) Asunción (Paraguay) Tomsk (Russia)

Number of people 49 97 16

Number of marks 2700 6100 1020

Number of restaurants to visit 67 254 32

Number of second activities (bars for aperitif, pubs or bars, clubs) 30 43 12

We analyzed this data and checked the following:

1) Co-rated activities number for different purposes. We considered number of users on the average co-rated 0, 1, 2, 3, 4, 5 activities (Tables 2 and 3).

2) Users similarity for different activities (Tables 4 and 5). We considered following similarity ranges: (0.6-0.8), (0.8-1.0)

Table 3

Co-rated activities number for Activity 2 (Visiting bars for aperitif, pubs or bars, clubs)

Number of Number of users on the average

co-rated Trento Asunción Tomsk

activities (Italy) (Paraguay) (Russia)

0 9 lO 2

1 ll 2O З

2 S lS l

3 5 9 l

4 З 2 2

5 З 2 2

Table 4

Users similarity for different activities, similarity range 0.6-0.8

EVALUATION OF ALGORITHMS

For each user u E U we created a dataset without all his ratings for the second activity a' (aperitif in Trento, bars in Asunción, club in Tomsk), thus defining Known(u, a') = 0. For each user u E U we built network of similar friends Net su (u) and network of similar users Nets f (u) based on user preferences for restaurants a. As a result and is a set of users

sharing similar preferences for a (e. g. dinner in a restaurant), since all user ratings for a' (e. g. drinking beer in a bar) were removed. Thus we are recommending places for , using the network of users with similar taste for a. We have used similarity measure with .

To evaluate performance of two approaches we have used the following definitions of precision and recall:

\\Tp(u)\\

PreClSl0n = \\TP(u)\\-\\FP(u)W

\\Tp(u)\\

reCaü \\Tp(u)\\-\\Fn(u)W

Tp{u) = {p £ Rec(u, a, g)\Liked(u,p, a, g)}, Fp(u) = {p £ Rec(u, a,g)\Disliked(u,p, a, g)}, Fn(u) = {p £ Unknown(u,a)

\Rec(u, a, g)\Liked(u, p, a, g)}.

Number of users on the average

Trento Asunción Tomsk

(Italy) (Paraguay) (Russia)

Activity 1 (Visiting restaurant) З 4 2

Activity 2 (Visiting bars for aperitif, pubs or bars, clubs) З 4 1

1,0 0,8 0,6 0,4 0,2 0,0

1

similar users

1 similar friends

Q

Number of users on the average

Trento Asunción Tomsk

(Italy) (Paraguay) (Russia)

Activity 1 (Visiting restaurant) 2 3 0

Activity 2 (Visiting bars for aperitif, pubs or bars, clubs) 2 4 0

similar users similar friends

Q

T

F

P

T

F

P

Table 5

Users similarity for different activities, Fig- 1 P^i^n in Kaly ^rentoX k = 1°

similarity range 0.8-1.0

Fig. 2. Recall in Italy (Trento), к = lO

similar users similar friends

Q

T

F

P

Fig. 3. Precision in Paraguay (Asuncion), k = 10

Fig. 4. Recall in Paraguay (Asuncion), k = 10

CONCLUSION

Analyzing the results we can see that in most cases precision of cross-domain recommendation using social factor (Facebook friends) appeared to be better in comparison with cross-domain recommendation using ^-nearest-neighbours approach (Fig. 1, 3, and 5). What is interesting to mention -is that recall in all cases is the best for the «Price/Quality» goal (Fig. 2, 4, and 6). And in case of Trento (Italy) in Fig. 2 and Tomsk (Russia) in Fig. 6 recall is also better for the «Bringing friends» goal. It is worth mentioning that for cross-domain recommendation it is important that we have high dense matrix for first domain, which we use in order to find users with similar preferences. As we see from our data analyses (Tables 2, 4, and 5) number of users on the average that corated more than four places is equal to one, mostly users co-rated two or three places, also there are not more than five, seven and two number of users on the average for Trento, Asunción and Tomsk correspondingly that has similarity higher than 0.6.

REFERENCES

1. Valeri B., Baez M., and Casati F., "Comealong: Empowering experience-sharing through social networks," in CollaborateCom, IEEE, 2012.

2. Valeri B., Baez M., and Casati F., "Come Along: understanding and motivating participation to social leisure activities," in CGC, IEEE, 2013.

3. Tou J. T., Gonzalez R. C., Pattern Recognition Principles. London - Amsterdam - Dom Mills, Ontario - Sydney -Tokyo: Addison-Wesley Publishing Company, 1974.

4. Segaran T., Programming Collective Intelligence. O'Reilly Media, Inc. Pub. Date: August 16, 2007. Print ISBN-13: 978-0-596-52932-1.

5. Jun Wang, Arjen P. de Vries, Marcel J. T. Reinders,

Unifying Userbased and Itembased Collaborative Filtering Approaches by Similarity Fusion, SIGIR'06, August 6-11, 2006, Seattle, Washington, USA. ACM 1-59593-369-7/06/0008.

6. Pazzani M. J., "A framework for collaborative, content-based and demographic filtering," ACM Transactions on Information Systems, vol. 22, no. 1, pp. 5-53, January 2004.

7. Xiaoyuan Su, Taghi M. Khoshgoftaar, "A survey of collaborative filtering techniques," Advances in Artificial Intelligence, vol. 2009 (2009), article ID 421425, 19 pages; doi:10.1155/2009/421425.

1,0

0,8 0,6 0,4 0,2 0,0

J J L

similar users similar friends

T F P Q

Fig. 5. Precision in Russia (Tomsk), k = all

1,2 1,0

similar users similar friends

similar users similar friends

T F P Q

ii i 111 il ■ ■ ■ 8

Fig. 6. Recall in Russia (Tomsk), k = all

ABOUT THE AUTHORS

BADRETDINOV, Marsel Rafaelevich, Department of Computer Science and Robotics, Ufa State Aviation Technical University (USATU), Ufa, Russia.

BADRETDINOV, Timur Rafaelevich, Department of Computer Science and Robotics, Ufa State Aviation Technical University (USATU), Ufa, Russia.

MAKEEV, Gregory Anatol'evich, Department of Computer Science and Robotics, Ufa State Aviation Technical University (USATU), Ufa, Russia.

CASATI, Fabio, Department of Information Engineering and Computer Science, University of Trento, Trento, Italy.

МЕТАДАННЫЕ

Заглавие: Повышение качества рекомендации мероприятий путем учета социального фактора в кросс-доменной рекомендации.

1 2 Авторы: М. Р. Бадретдинов , Т. Р. Бадретдинов , Г. А. Макеев3, Ф. Касати4 Организации: 1-3 ФГБОУ ВПО «Уфимский государственный авиационный технический университет» (УГАТУ), Россия.

4

Университет Тренто, Италия. Email: 1 timars@list.ru. Язык: английский.

Источник: Вестник УГАТУ. 2014. Т. 18, № 5 (66). С. 91-95,

ISSN 2225-2789 (Online), ISSN 1992-6502 (Print). Аннотация: Оценено влияние социальных факторов таких как дружба в Facebook, на качество кросс-доменной рекомендации. Для этого были проанализированы предпочтения людей из нескольких городов в разных странах по поводу различных мероприятий (досуга). При этом учитывались цели с которыми посещались мероприятия (проводился досуг). Ключевые слова: рекомендательные системы; совместная фильтрация; социальная рекомендация; кросс-доменная рекомендация.

Об авторах:

БАДРЕТДИНОВ Марсель Рафаэлевич, асп. каф. выч. мат и кибернетики. Дипл. мат.-программист (УГАТУ, 2012).

БАДРЕТДИНОВ Тимур Рафаэлевич, асп. каф. выч. мат и кибернетики. Дипл. мат.-программист (УГАТУ, 2012).

МАКЕЕВ Григорий Анатольевич, доц. каф. выч. мат. и кибернетики. Дипл. инж.-прогр. (УГАТУ, 2003). Канд. техн. наук (УГАТУ, 2006).

КАСАТИ Фабио, профессор, факультет информационной инженерии и информатики. PhD (Миланский поли-техн. ун-т, 1999). Иссл. в обл. соц. информатики.