Extending Content-Boosted Collaborative Filtering for Context-aware,

Mobile Event Recommendations

Daniel Herzog and Wolfgang W

orndl

Department of Informatics, Technische Universit

at M

unchen, Boltzmannstr. 3, 85748 Garching, Germany

Keywords:

Recommender System, Event Recommendations, Content-Boosted Collaborative Filtering, Context-

awareness, Mobile Application.

Abstract:

Recommender systems support users in ﬁltering large amounts of data to ﬁnd interesting items like restaurants,

movies or events. Recommending events poses a bigger challenge than recommending items of many other

domains. Events are often unique and have an expiration date. Ratings are usually not available before the

event date and not relevant after the event has taken place. Content-boosted Collaborative Filtering (CBCF)

is a hybrid recommendation technique which promises better recommendations than a pure content-based or

collaborative ﬁltering approach. In this paper, CBCF is adapted to event recommendations and extended by

context-aware recommendations. For evaluation purposes, this algorithm is implemented in a real working

Android application we developed. The results of a two-week ﬁeld study show that the algorithm delivers

promising results. The recommendations are sufﬁciently diversiﬁed and users are happy about the fact that

the system is context-aware. However, the study exposed that further event attributes should be considered as

context factors in order to increase the quality of the recommendations.

1 INTRODUCTION

Recommender systems are information ﬁltering and

decision support tools providing personalized recom-

mendations by identifying information or products

which best satisfy the user’s needs. Recommender

systems are increasingly used in a mobile context

due to the widespread use of smartphones and tablets.

These devices allow more accurate recommendations

because they can identify the context of the recom-

mendation in a more detailed manner (Ricci, 2011).

In general, context can be described as ”any informa-

tion that can be used to characterize the situation of

entities (i.e., whether a person, place, or object) that

are considered relevant to the interaction between a

user and an application, including the user and the

application themselves” (Dey et al., 2001). Mobile

devices are able to collect this type of information

as they are often equipped with sensors which allow

them to identify context data like the user’s current

position or speed of travel.

Event recommendations pose a new challenge in

the ﬁeld of recommender systems. Minkov et al.

(2010) explain that events, as opposed to movies or

restaurants, usually take place only one-time under

the same conditions, thus they come with an expi-

ration date. User ratings, which can be considered

for recommendations, are usually not available be-

fore the event takes place and no longer of importance

when the event is over or expired. This lack of ratings

makes it impossible to recommend, for example, fu-

ture events other users with similar preferences have

liked. Hence, additional data has to be collected and

processed by more sophisticated techniques in order

to generate recommendations for events. Event rec-

ommendations are moreover a good example show-

ing the importance of context in recommender sys-

tems. For instance, in case of a bad weather forecast,

recommendations for outdoor events could be inap-

propriate.

Different techniques are available for recommend-

ing items such as events. Hybrid recommenders use

the combination of two or more techniques in order

to overcome weaknesses of single techniques and to

improve the quality of recommendation. Melville

et al. (2002) present Content-boosted Collaborative

Filtering (CBCF), a hybrid recommendation tech-

nique which generates better recommendations than

a pure content-based (CB) or a pure collaborative ﬁl-

tering (CF) approach by combining these two tech-

niques.

The main goal of this work is to identify new

methods to improve the quality of event recommenda-

Herzog, D. and Wörndl, W.

Extending Content-Boosted Collaborative Filtering for Context-aware, Mobile Event Recommendations.

In Proceedings of the 12th International Conference on Web Information Systems and Technologies (WEBIST 2016) - Volume 2, pages 293-303

ISBN: 978-989-758-186-1

293

tions in a mobile scenario. For this purpose, we intro-

duce a new approach to generate personalized event

recommendations. To the best of our knowledge, we

are the ﬁrst to use CBCF within a context-aware rec-

ommender to recommend all kinds of cultural events.

As a ﬁrst step, this paper presents important recom-

mendation techniques and CBCF is explained. We

further explain the details of our algorithm and show

how it can be used for recommending events. Our

Android application which implements the context-

aware CBCF algorithm is then presented and used for

the ﬁeld study of the system. After discussing our

ﬁndings, we present related work in the ﬁeld of rec-

ommender systems and event recommendations. The

paper ﬁnishes with a conclusion and an outlook on

future work in this area.

2 RECOMMENDATION

TECHNIQUES

Different techniques and algorithms can be imple-

mented in recommender systems in order to gener-

ate personalized recommendations. These techniques

differ in the information they collect and how they

rate items. Two popular techniques are CB predic-

tions and CF.

CB systems try to recommend items which are

similar to those the user has liked in the past (Bala-

banovi

c and Shoham, 1997). A profound knowledge

of the item representation is mandatory because CB

recommender systems analyze item descriptions and

attributes to identify similar items. Pazzani and Bill-

sus (2007) differentiate between structured and un-

structured data to represent items: items represented

by structured data comprise a set of attributes and

there is a known set of values that each of these at-

tributes may have. Unstructured data have no well-

deﬁned values, e.g., a text ﬁeld allowing entry into ev-

ery possible text. In most cases, semi-structured data,

a combination of attributes with a known set of val-

ues and free-text ﬁelds, is chosen. One advantage of

CB recommending is that no critical mass of users is

necessary to provide recommendations. Nevertheless,

CB recommending comes with some limitations. As

the New User Problem explains, a new user has to rate

a signiﬁcant number of items before the system can

offer accurate recommendations for her or him. Fur-

thermore, CB systems aim to recommend items which

tightly ﬁt a user’s proﬁle, thus a lack of diversity can

be an issue (Adomavicius and Tuzhilin, 2005).

Case-based recommendations are a special case

of CB recommenders. According to Smyth (2007),

cases are solutions to a given recommendation prob-

lem. The simplest scenario is the recommendation of

the top k most similar cases matching a user query.

Case-based recommendations can be used for recom-

mending items with a structured item representation

as they allow the use of sophisticated approaches to

judge how cases respond to the user query. Formula 1

is one exemplary approach how to calculate the simi-

larity between a case and the query:

Similarity(q, c) =

∑

i=1

· sim

, c

)

∑

i=1

, (1)

(Smyth, 2007), where n is the number of attributes

of the item. sim

is the similarity between an attribute

i of a query q and a case c. The attribute similarity

can be calculated with formula 2 and is weighted by

sim

, c

) = 1 −

− c

max(q

, c

)

(2)

(Smyth, 2007). The presented formulas depend on

the selected similarity metric. Smyth differentiates

between symmetric and asymmetric similarity met-

rics. A symmetric similarity metric reduces the sim-

ilarity by the same value if the case attribute value

is lower or higher than the query attribute value. An

asymmetric metric prefers either higher or lower val-

ues.

In contrast, CF recommends items other users

with similar preferences have liked (Balabanovi

c and

Shoham, 1997). In the majority of cases, nearest

neighbor algorithms are implemented in CF recom-

menders. Schafer et al. (2007) differentiate between

User-Based Nearest Neighbor and Item-Based Near-

est Neighbor algorithms. While User-Based Nearest

Neighbor algorithms call users who rate objects sim-

ilar neighbors, Item-Based Nearest Neighbor algo-

rithms rate items based on similarities between items

(Sarwar et al., 2001). Like CB systems, CF has some

limitations. In addition to the New User Problem,

the New Item Problem is an issue because items not

rated by a substantial number of users cannot be rec-

ommended (Adomavicius and Tuzhilin, 2005). These

problems deﬁne the Cold-Start Problem - a serious

problem for event recommendations. Because users

cannot rate unique future events before actually at-

tending them, most of the items in the system remain

unrated (Dooms et al., 2011). The limitation of such a

sparely ﬁlled user-item rating matrix is called sparsity

(Melville et al., 2002). Sparsity causes a low proba-

bility of ﬁnding a set of users with signiﬁcantly simi-

lar ratings, thus leading to fewer or no recommenda-

tions.

The presented limitations of CB and CF recom-

menders can be illustrated using a simple example.

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

294

Table 1 shows a user-item rating matrix with two

users A and B and ﬁve events. Four events are as-

signed to the genre comedy, thus they are assumed to

be similar. The other event is a musical. User A al-

ready provided a good rating to two comedy events,

indicated by the + symbol. User B gave positive rat-

ings to the two other comedy events and to the musi-

cal. A CB recommender system is able to recommend

to each user the comedy events not previously rated

as these events are similar. As CB recommendation is

not taking other users into account, no further recom-

mendations for User A are possible. A CF approach,

however, is not able to recommend any event to any

user. User A and B had not initially rated any com-

mon events, thus they cannot be identiﬁed as neigh-

bors.

Table 1: Exemplary user-item rating matrix for events.

User A User B

Comedy 1 +

Comedy 2 +

Comedy 3 +

Comedy 4 +

Musical 1 +

Hybrid recommender systems combine two or

more techniques in order to overcome such limi-

tations. One example of a hybrid recommending

method is CBCF. CBCF is a feature augmentation

hybrid which uses the output from a CB prediction

to generate recommendations in the subsequent CF

phase (Burke, 2002). The initial CB prediction is ex-

ecuted on the user-item rating matrix containing all

ratings given by the users. The predicted ratings are

then stored in the user-item rating matrix, now called

pseudo user-item rating matrix, which is character-

ized by a lower sparsity. Finally, the CF algorithm is

executed on the pseudo matrix (Melville et al., 2002).

In the presented example (Table 1), the CB approach

extends the user-item rating matrix by CB predictions,

in this case, two comedy events for each user. Based

on the new pseudo matrix, CF is able to identify User

A and B as neighbors. This allows the recommenda-

tion of the musical to User A since it was highly rated

by a user with similar preferences.

3 CONTEXT-AWARE

CONTENT-BOOSTED

COLLABORATIVE FILTERING

FOR MOBILE EVENT

RECOMMENDATIONS

The example presented in Table 1 illustrates the

strengths of hybrid recommenders. CB recom-

menders identify events similar to those the user al-

ready liked. To arrive at this conclusion, past feed-

back has to be analyzed but no critical mass of users

is necessary. CF promises an increase in the variety

of the recommendations but because of the lack of

ratings for future events, CF has to be combined with

other techniques. This is why we use CBCF for our

event recommender as it combines these two tech-

niques and promises better recommendations than

pure CB or CF recommenders (Melville et al., 2002).

Furthermore, we want to show that CBCF can be used

for context-aware recommendations. We brieﬂy intro-

duced the idea of our algorithm and ﬁrst results of a

ﬁeld study in (Herzog and Woerndl, 2015).

Adomavicius and Tuzhilin (2011) present three

different paradigms for incorporating contextual in-

formation into a recommendation process. Our sug-

gested context-aware CBCF approach implements

contextual pre-ﬁltering to diminish the amount of

events available for recommendations before the ac-

tual recommendation process takes place. This ap-

proach is advantageous because events which are im-

possible to recommend are excluded immediately: for

example, an event taking place too far away from the

user’s location. In future work, we plan to try the

other paradigms for incorporating context as well and

compare those results to the solution presented in this

paper.

After the pre-ﬁltering phase, our algorithm ana-

lyzes user feedback on past recommendations in or-

der to predict the ratings of the remaining events in

the CB recommendation phase. These predictions are

entered in the pseudo rating matrix. Finally, the CF

phase is executed in order to consider all predicted

ratings of all users and identify additional recommen-

dations. This step is necessary to increase the vari-

ety of our event recommendations. To sum up, the

proposed recommendation algorithm comprises three

phases:

1. Contextual pre-ﬁltering

2. Content-based prediction

3. Collaborative ﬁltering

In the following, we describe these phases in detail.

Extending Content-Boosted Collaborative Filtering for Context-aware, Mobile Event Recommendations

295

3.1 Contextual Pre-ﬁltering

Before predicting ratings for the events available in

the database, the number of possible recommenda-

tions is decreased by taking context into account. Rel-

evant context factors in this work were identiﬁed in

expert interviews with selected representatives from

the German event and ticketing industry. These con-

text factors are:

• Current Position: It is likely that the user prefers

selected venues but the travel distance to these

venues has to be appropriate. The system should

be able to identify the user’s current location and

the user should be able to set a radius around it.

Only venues within this radius should be consid-

ered for recommendation.

• Temporary Preference of Selected Genres: The al-

gorithm should ignore certain genres during the

recommendation process, e.g., when a genre is in-

appropriate for the user’s companions. The user

should be able to select or deselect genres in or-

der to tell the system the appropriate genres.

• Budget: The algorithm has to respect the user’s

available budget. For this purpose, the user has to

set an upper limit for event prices or the sum she

or he can spend per week or month.

• Weekday: Recommendations have to respect the

identiﬁed days the user is available for events

(e.g., only on weekends). The user should be

able to tell the system the weekdays for the pre-

ﬁltering of events.

• Time of Day: Recommendations are only useful

when the user is available at the suggested time

of day (e.g., not during the morning). The user

should be able to tell the system the times of day

for the pre-ﬁltering of events.

• Scheduling Conﬂict: If the user already purchased

tickets for a recommended event, no further events

which take place at the same time should be rec-

ommended. The recommender system has to

identify such conﬂicts automatically and exclude

events if necessary.

As described, these factors are used as criteria for

exclusion. If the context of a potential recommenda-

tion exceeds a deﬁned threshold, e.g., the distance to

the venue, the corresponding event will not be consid-

ered for recommendation. The developed prototype

in section 4 allows the user to set and modify these

thresholds as explained above.

3.2 Content-Based Prediction

After excluding events which do not satisfy the con-

text constraints, the classical CBCF approach can be

executed. At ﬁrst, the CB prediction phase of CBCF

has to be adapted to the special case of event recom-

mendations. In this section, we describe how we an-

alyze event attributes rated by the user in the past in

order to estimate ratings for events comprising these

attributes. At this point, context is no longer the fo-

cus. Nevertheless, the presented context factors are

reﬂected in the item representation during the CB pre-

diction, e.g., when determining how much the user

likes a certain venue.

As explained, items can be represented by struc-

tured data, unstructured data and semi-structured

data. Based on the expert interviews we conducted,

events are mainly characterized by structured data. A

Munich-based event and ticketing company provided

a dataset with approximately 3700 real events which

were used for the survey in section 5. The dataset in-

cludes the following, relevant event attributes:

• The event name

• The name and address of the venue

• The genre

• The exact date when the event starts

• The vendor

The dataset did not provide information about the

ticket price which therefore was not considered in this

work. The structured characteristic of events is the

reason why we propose a case-based approach for the

CB prediction of ratings.

In order to adapt CBCF to event recommenda-

tions, we propose a slightly different similarity metric

than the symmetric and asymmetric similarity metrics

presented in section 2. Nominal case attributes such

as the genre or the venue are treated as binary values.

If an event takes place at a certain venue, its value is

1 for this venue and 0 for all other venues. The query

attributes depend on the user’s history. If a user liked

90% of all recommendations of a venue, the query

value q

for this venue is 0.9. The attribute similar-

ity is calculated with Formula 2 using q

equals 0.9

and the case value c

equals 0 or 1 as input. Other

attributes, such as the event’s price, could apply an

asymmetric similarity metric. If the price is lower

than the average as identiﬁed by the user’s history, the

similarity will be less reduced than for a higher price.

The challenge is to ﬁnd a way to calculate the

query value q

for each attribute. As described, user

feedback for an upcoming event is usually not avail-

able up-front, hence the user’s history has to be used

as a basis for the calculation. For event recommenda-

tions, two scenarios are possible: the user either likes

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

296

or rejects a recommendation. Additionally, if the user

actually purchases tickets for a recommended event,

this would be considered a more positive feedback

than just liking the recommendation. In this work, we

count all liked, rejected and purchased recommenda-

tions and calculate the share of the positive feedback.

Thereby, liked events are increased by factor 3 and

events with purchased tickets by factor 5 as we as-

sume that users often have to reject recommendations

only because of time constraint issues. If a user liked

one theater recommendation, rejected another theater

recommendation but purchased tickets for a third the-

ater performance, q

for i = theater will be calculated

≈ 0.8889. This value means that the user likes

the attribute theater in a recommendation with a prob-

ability of 88.89%.

If q

can be calculated for a sufﬁcient number of

attributes, i.e., past feedback is available for these at-

tributes, a prediction is called accurate. Formula 1 can

be used to calculate the similarity between the item

and the query corresponding to the value of the rec-

ommended item for the user. This value is stored in

the pseudo user-item rating matrix and used for the

upcoming CF phase.

3.3 Collaborative Filtering

Recommendations which were not rated during the

CB phase are candidates for the upcoming CF phase.

Rating recommendations in the CF phase is of prime

importance because focusing solely on the user his-

tory could lead to a poor diversity of recommen-

dations. As explained, User-Based Nearest Neigh-

bor and Item-Based Nearest Neighbor algorithms are

available for CF. In this work, we implement a User-

Based Nearest Neighbor algorithm as it is already

used for the CBCF approach in (Melville et al., 2002).

Some assumptions have to be made to reach valu-

able predictions. Only users with a similarity of at

least 50% are considered as neighbors. The CF ﬁlls

the pseudo matrix with additional ratings. In the end,

the events with the highest value for the user are rec-

ommended. Every recommendation has to have a

value of at least 0.5 on a scale from 0 to 1. In or-

der to achieve a high serendipity, we divide the events

in three groups. Around one third of the recommen-

dations are events which take place within 7 days,

one third within one month and one third not within

one month. If not enough events are available for a

group, the list of recommendations is ﬁlled with the

best available recommendations. We limit the max-

imum number of recommendations at one request to

10 to avoid overwhelming the user with events. As

a result, our context-aware CBCF method is able to

recommend a set of events which respect the user’s

context and expectations and promise a satisfying di-

versity.

4 DEVELOPED MOBILE

PROTOTYPE

We developed an Android application to implement

our algorithm and to evaluate the recommender sys-

tem. The application can be installed on devices run-

ning Android 4 or newer.

For our prototype, the dataset is stored in a

MySQL database connected to a RESTful web ser-

vice we developed. When a client demands recom-

mendations from the web service, it transmits the user

id and the current context information to the server.

The web service then is able to provide personalized

recommendations based on the user-item rating ma-

trix which is stored centrally in the MySQL database.

These recommendations are received by the applica-

tion and displayed to the user.

In this section, we brieﬂy present the applica-

tion. The application can be used without a login. A

unique, device-dependent id allows personalized rec-

ommendations without creating a user account. After

starting the app, a request is sent to the server auto-

matically and the user immediately receives a list of

new recommendations (Figure 1). Recommendations

are presented as cards, a design principal made pop-

ular through applications such as Tinder and Google

Now. The main advantages of this representation are

the low cognitive load because of the small number

of items visible and the fact that the user can navigate

through the app and provide feedback with only one

hand (Torkington, 2014). In the developed app, the

cards provide the most interesting information about

the recommended event such as its name, the genre,

the location and the date. Furthermore, the calculated

rating for the user is expressed as a percentage and

presented together with a short explanation. For ex-

ample, the explanation ”dein Feedback” (German for

”your feedback”) indicates that the event is recom-

mended because of a rating from the CB recommen-

dation phase. The user can swipe recommendations to

the right to give a positive rating to a recommendation

and to the left in order to reject the recommendation

and to provide a negative rating.

Clicking on a recommendation displays a detailed

view of the selected event. At the selected event, the

user can ﬁnd additional information concerning the

venue or ﬁnd Facebook friends who are attending the

event. A button linking to the ticket vendor is also

available. For testing purposes, the prototype assumes

Extending Content-Boosted Collaborative Filtering for Context-aware, Mobile Event Recommendations

297

Figure 1: New recommendations provided by the recom-

mender system.

that a user clicking this button eventually purchases a

ticket.

Using the Navigation Drawer, the app’s menu bar,

the app user can call up the settings view allowing her

or him to modify the thresholds for the contextual pre-

ﬁltering (cf. section 3.1). Figure 2 shows a screen-

shot of the geographical context settings. The user

can choose the radius threshold for the contextual pre-

ﬁltering of events by moving the slider thumb. The

radius is drawn around a certain position which can

be either determined by activating the device’s GPS

sensor or chosen from a predeﬁned list with cities. In

this example, the current user position is set to the

city center of Munich and only events within a ra-

dius of 25 kilometers are considered for recommen-

dation. In the second settings view, the user can de-

termine her or his desired weekdays and times of day

for the pre-ﬁltering phase by selecting the correspond-

ing time slots (Figure 3). For this purpose, we offer

a calendar view which splits every day of the week

into ﬁve times of day from morning (6 am until 11

am) until night (10 pm until 6 am). This design al-

lows the user to select time slots for each day individ-

ually. This selection option is important for event rec-

ommendations because we believe a potential visitor

cannot attend events every day at the same time. The

additional check box allows users to select and dese-

lect all slots at once. The third setting view displays a

list of genres which can be selected or deselected by

the user in order to determine appropriate genres for

the pre-ﬁltering phase.

Figure 2: Geographical context settings.

Figure 3: Temporal context settings.

5 FIELD STUDY

In this section, we describe a ﬁeld study conducted to

evaluate the developed recommender system and the

context-aware CBCF algorithm. The study was con-

ducted after the implementation of the presented sys-

tem. The main goal of the study was to test the rec-

ommender system in a realistic scenario. The study

was intended to deliver insights into the quality of the

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

298

recommender system. The method of inﬂuencing the

context-awareness in the app settings also had to be

evaluated. The study results are presented and inter-

preted at the end of this section.

5.1 Setting and Procedure

The ﬁeld study was conducted as a two-week beta

test, meaning, the users installed the application on

their own smartphones. Users were advised to use the

application as if they were using an application they

installed voluntarily in everyday life. This means that

during the test, users were allowed to use the system

whenever and wherever they wanted to, in the desired

intensity. Nevertheless, they were recommended to

use the app at least for a certain amount of time to

get a ﬁrst impression of the system and the delivered

recommendations.

After two weeks, the beta test terminated and a

survey was sent to the participants. The main objec-

tive of the survey was to evaluate: the algorithm, the

offered method to inﬂuence the context-awareness,

the user interfaces and the recommender system as a

whole. Table 2 lists all evaluation statements included

in the survey. The participants had to rate these state-

ments using a 5-point Likert scale with 1 represent-

ing no agreement at all and 5 representing complete

agreement.

Table 2: Survey statements.

# Statement

1 Overall, I like the recommender system

2 The recommendations meet my expecta-

tions

3 The recommendations are sufﬁciently di-

versiﬁed

4 My feedback leads to better recommenda-

tions

5 The system provides sufﬁcient means to

express my expectations

6 I like the way I can set local constrains

7 I like the way I can set temporal constrains

8 I like the way I can choose certain genres

9 The user interfaces are intuitive

10 I like the presentation of new recommen-

dations

Further personal questions were added to the sur-

vey in order to obtain an overview of the participant’s

background, personal experience with events and per-

sonal experience with similar recommender systems.

5.2 Participants

The participants were selected to achieve a cross-

sectional survey of event visitors. Different age

groups were considered as were casual visitors and

so-called expert visitors who attend events on a regu-

lar basis. Even though the participants were obligated

to own a smartphone, less technophilic participants

were as important as more experienced users.

Twenty-one participants started the ﬁeld study of

whom 16 terminated successfully after two weeks by

completing the survey. Of the 16 participants who

completed the study, one was younger than 18 years

old, three were between the ages of 18 and 25, nine

were between 26 and 35, two were between 36 and

45 and one was older than 45 years old. The aver-

age technical afﬁnity of the participants is 3.73 (σ:

0.88) on a 5-point Likert scale according to their own

estimation. Mobile event applications were used by

62.5% of the participants once a month or less often

and by 18.8% at least once a week. The participants

reported different experiences with recommender sys-

tems such as the Amazon website. Recommender

systems were used several times per week by 25% and

43.8% used them once per month or less often. About

a third of the participants, 37.5%, can be called expert

visitors as they attend events requiring a ticket pur-

chase once per month or more often whereas 62.5%

purchase tickets not more often than a few times per

year.

5.3 Results

The responses (Figure 4) show that the participants

are satisﬁed with our solution (∅: 3.75, σ: 0.83).

Only 25% of the participants rated the system with a

3 or less. We also wanted to know if the received rec-

ommendations met their expectations. The responses

are slightly above average (∅: 3.38, σ: 0.60). Ac-

cording to the survey results, the recommendations

can be called sufﬁciently diversiﬁed (∅: 3.63, σ:

1.05). The participants believe that their feedback im-

proves the personalized recommendations (∅: 3.69,

Figure 4: Overall satisfaction with the recommender sys-

tem.

Extending Content-Boosted Collaborative Filtering for Context-aware, Mobile Event Recommendations

299

σ: 0.77). The frequency of usage differed between

participants. The 16 participants would have bought

a total of 16 tickets based on the recommendations

which means one sold ticket per person during the

two week study period. A majority of the participants,

87.5%, mentioned that they would like to continue us-

ing the system to ﬁnd interesting events in the future.

In general, the users were satisﬁed with the choice

of settings which allowed them to modify the context-

awareness (∅: 3.5, σ: 0.79). They particularly liked

how they were able to inﬂuence the geographical con-

text (∅: 3.94, σ: 0.75). A higher variance of ratings

can be observed when asking the participants about

the calendar view (∅: 3.56, σ: 1.17) and the list of

genres (∅: 3.44, σ: 1.12). A free text ﬁeld at the

end of the survey allowed the participants to add fur-

ther thoughts. Some participants requested further op-

tions to modify the context-awareness, for example,

the user’s current budget for events.

The developed Android application is only one ex-

ample of how our algorithm can be implemented in a

mobile event recommender system. According to the

participants, the apps’ user interfaces are intuitive (∅:

4.06, σ: 0.97) and they were pleased with the chosen

card layout for the new recommendations (∅: 4.07,

σ: 0.77).

5.4 Discussion

Our ﬁeld study allowed us to evaluate the recommen-

dations generated by our context-aware CBCF algo-

rithm in a realistic scenario. The results show that

our approach to adapt CBCF is a promising way to

generate personalized event recommendations. The

user’s feedback on past recommendations is a valu-

able basis to ﬁll the user-item rating matrix which

can be extended by a CB prediction. The CF algo-

rithm based on this matrix is able to provide accurate

recommendations and ensures a sufﬁcient diversity of

events. Furthermore, our system is able to provide

recommendations even if only one user is using the

system.

CBCF can be extended by context-aware recom-

mendations, another important ﬁnding of the study.

The recommendations do not only satisfy the search

activities of the users, but the participants in our sur-

vey were also pleased they were able to inﬂuence the

context-awareness. Nevertheless, our study reveals

that further context factors should be identiﬁed and

considered during the recommendation process. Our

ﬁrst prototype focuses on the most obvious context

factors such as the user’s current location or tempo-

ral context factors. In addition to the price of the

event, further context factors which were completely

ignored during this work should be incorporated in

order to improve the quality of the recommendations

in different situations. Examples of such factors are

the current weather or if the user is alone or accom-

panied by someone. In this case, a group recom-

mender could help to ﬁnd suitable recommendations

for all group members. If, for example, a fan of music

events is accompanied by her or his children, recom-

mending a child-friendly event will be more appro-

priate than a visit to a concert. While our ﬁrst pro-

totype mainly supports users in receiving recommen-

dations at a certain time before the actual event takes

place, we believe that additional, contextual informa-

tion would also allow to ﬁnd suitable recommenda-

tions even more spontaneously (Herzog and Woerndl,

2015).

According to the feedback we got from our par-

ticipants, we learned that the users of mobile rec-

ommender systems value appealing user interfaces.

The chosen card layout for new recommendations re-

ceived positive feedback, conﬁrming our belief that

this layout is an appropriate way to present recom-

mendations and to offer a quick and easy solution for

providing feedback at the same time.

6 RELATED WORK

In comparison to movie recommendations, research

has paid little attention to recommender systems for

events. Minkov et al. (2010) present an approach

which considers the individuals’ preferences for past

events and combines these preferences with other

peoples’ likes and dislikes. Dooms et al. (2011) con-

ducted a user-centric evaluation of different event rec-

ommender algorithms. Results show that a hybrid

of content-based and collaborative ﬁltering performs

better than other algorithms in most quality factors

like accuracy, satisfaction and usefulness. The hy-

brid approach of Khrouf and Troncy (2013) combines

content-based and collaborative ﬁltering and is en-

riched by Linked Data to overcome data sparsity. Cor-

nelis et al. (2005) model user and item similarities

as fuzzy relations in their hybrid approach. Outlife

is an event recommender which groups friends in so-

cial groups (De Pessemier et al., 2013). It allows the

creation of recommendations for speciﬁc groups of

friends and can recommend friends who should be in-

vited to an event. Zhang et al. (2013) present a group

recommender which recommends event-based groups

using matrix factorization. A study using Meetup.com

data shows the effectiveness of their approach. Quer-

cia et al. (2010) analyzed mobile phone location data

to understand how the user’s current location inﬂu-

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

300

ences the acceptances of social event recommenda-

tions. They found out that recommending events that

are popular among residents of an area is more bene-

ﬁcial than recommending nearby events. This ﬁnding

can be seen as a promising solution for the Cold-Start

Problem in event recommender system. If the event

is location independent, Daly and Geyer (2011) sug-

gest using the popularity as a metric for overcoming

the Cold-Start Problem. A special type of events that

can be recommended are distributed events. These

events are collections of smaller, single but very sim-

ilar events that occur at the same day (Schaller et al.,

2013).

A few mobile applications offering personalized

event recommendations are already on the market.

One example is Bandsintown, an application available

for Android and iOS devices which focuses on music

events (Figure 5). Recommendations made for events

take into consideration music selections locally stored

on the user’s device. Furthermore, external sources

like Facebook or Twitter can be connected to the ap-

plication in order to collect additional information

about user preferences in regard to music. Other ex-

amples of mobile applications which offer some kinds

of event recommendations are the EVENTIM DE ap-

plication, offered by the German ticketing and event

company CTS Eventim and XING EVENTS, an ap-

plication of the German based social network for pro-

fessionals XING. The fact that the ticketing and event

industry has already started to implement personal-

ized recommendations in their offers underlines the

signiﬁcance of the topic for the different players in-

volved in this industry. Visitors, on the one hand, can

ﬁnd interesting events nearby which they might over-

look otherwise. On the other hand, the recommen-

dations promise additional sales and the users’ feed-

back allows organizers to predict the number of fu-

ture visitors for similar events (Minkov et al., 2010).

Setten et al. (2004) focus on context-aware sys-

tems. The mobile application COMPASS demon-

strates how user interests and context can be respected

in a recommender systems. Oku et al. (2006) show

that recommendations which take the user’s context

into consideration lead to a higher user satisfaction. A

general model for context-aware and proactive recom-

mending is delivered by Woerndl et al. (2011). Ado-

mavicius and Tuzhilin (2011) emphasize the impor-

tance of context in recommender systems and intro-

duce three different paradigms for incorporating con-

textual information into the recommendation process.

The related work shows the potential of event rec-

ommender systems and the advantages of context-

awareness. But to the best of our knowledge, there

is no work examining the use of CBCF within a

Figure 5: Bandsintown Android application.

context-aware recommender to recommend all kinds

of events. We closed this gap by showing that CBCF

can be extended by context-aware recommendations

and adapted to the scenario of event recommenda-

tions. Our results show that the chosen approach is

able to deliver promising recommendations.

7 CONCLUSION AND FUTURE

WORK

In this paper, we introduced a new method for per-

sonalized event recommendations. We developed an

algorithm based on CBCF and extended it by context-

aware recommendations. This algorithm was imple-

mented in an Android application and evaluated in a

two-week ﬁeld study. According to the results, our

approach delivers promising recommendations. By

combining different recommendation techniques, cer-

tain limitations, such as the New Item Problem and a

lack of diversity, can be overcome. We believe that

the results of the ﬁeld study emphasize the economic

potential of the developed system and underline the

importance of mobile recommender systems in the

ticketing and event industry. The ﬁrst applications of

recommender systems that provide some kind of per-

sonalized recommendations are already available and

are steadily gaining popularity.

Future work should focus on identifying addi-

tional context factors to be considered during the al-

gorithm execution. These factors should be capable

Extending Content-Boosted Collaborative Filtering for Context-aware, Mobile Event Recommendations

301

of ﬁnding suitable recommendations in different situ-

ations, e.g., for tourists who spend only a very short

time visiting a city. Improvements are also possible

in regards to other solutions of context incorporation.

Currently, we use pre-ﬁltering only, but post-ﬁltering

and contextual modeling should be considered as well

in future work. The contextual pre-ﬁltering immedi-

ately excludes inappropriate events but we should ex-

amine if users are willing to accept little violations of

the thresholds if an event really satisﬁes their needs.

Furthermore, social recommendations promise an im-

provement of the recommendations. Events of par-

ticular importance for the user, e.g., events the user

liked on social networks like Facebook, could receive

a bonus on the calculated rating to increase the prob-

ability of recommendation. Another aspect is implicit

user feedback which was not collected by this ﬁrst

prototype but could be used to improve the recom-

mendations. Our ﬁrst prototype or pure CB or pure

CF approaches could be used as a baseline application

in order to evaluate the suggested extensions and their

inﬂuence on the quality of the recommendations. For

this purpose, we want to conduct another study with a

larger number of users which allows us to obtain more

in-depth user feedback and statistics.

ACKNOWLEDGEMENTS

This work is part of the TUM Living Lab Connected

Mobility (TUM LLCM) project and has been funded

by the Bayerisches Staatsministerium f

ur Wirtschaft

und Medien, Energie und Technologie (StMWi).

REFERENCES

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

generation of recommender systems: A survey of the

state-of-the-art and possible extensions. IEEE Trans.

on Knowl. and Data Eng., 17(6):734–749.

Adomavicius, G. and Tuzhilin, A. (2011). Context-aware

recommender systems. In Ricci, F., Rokach, L.,

Shapira, B., and Kantor, P. B., editors, Recommender

Systems Handbook, pages 217–253. Springer US.

Balabanovi

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

based, collaborative recommendation. Commun.

ACM, 40(3):66–72.

Burke, R. (2002). Hybrid recommender systems: Survey

and experiments. User Modeling and User-Adapted

Interaction, 12(4):331–370.

Cornelis, C., Guo, X., Lu, J., and Zhang, G. (2005). A

fuzzy relational approach to event recommendation.

In Prasad, B., editor, IICAI, pages 2231–2242. IICAI.

Daly, E. M. and Geyer, W. (2011). Effective event discov-

ery: Using location and social information for scoping

event recommendations. In Proceedings of the Fifth

ACM Conference on Recommender Systems, RecSys

’11, pages 277–280, New York, NY, USA. ACM.

De Pessemier, T., Minnaert, J., Vanhecke, K., Dooms, S.,

and Martens, L.(2013). Social recommendations for

events. In 5th ACM RecSys Workshop on Recom-

mender Systems and the Social Web.

Dey, A. K., Abowd, G. D., and Salber, D. (2001). A

conceptual framework and a toolkit for supporting

the rapid prototyping of context-aware applications.

Hum.-Comput. Interact., 16(2):97–166.

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

A user-centric evaluation of recommender algorithms

for an event recommendation system. In Proceed-

ings of the RecSys 2011 : Workshop on Human

Decision Making in Recommender Systems (Deci-

sions@RecSys’11) and User-Centric Evaluation of

Recommender Systems and Their Interfaces - 2

(UCERSTI 2) afﬁliated with the 5th ACM Conference

on Recommender Systems (RecSys 2011), pages 67–

73.

Herzog, D. and Woerndl, W. (2015). Spontaneous event

recommendations on the go. In Proceedings of the

2nd International Workshop on Decision Making and

Recommender Systems (DMRS2015), Bolzano, Italy,

October 22-23, 2015.

Khrouf, H. and Troncy, R. (2013). Hybrid event recommen-

dation using linked data and user diversity. In Pro-

ceedings of the 7th ACM Conference on Recommender

Systems, RecSys ’13, pages 185–192, New York, NY,

USA. ACM.

Melville, P., Mooney, R. J., and Nagarajan, R. (2002).

Content-boosted collaborative ﬁltering for improved

recommendations. In Eighteenth National Conference

on Artiﬁcial Intelligence, pages 187–192, Menlo Park,

CA, USA. American Association for Artiﬁcial Intelli-

gence.

Minkov, E., Charrow, B., Ledlie, J., Teller, S., and Jaakkola,

T. (2010). Collaborative future event recommenda-

tion. In Proceedings of the 19th ACM International

Conference on Information and Knowledge Manage-

ment, CIKM ’10, pages 819–828, New York, NY,

USA. ACM.

Oku, K., Nakajima, S., Miyazaki, J., and Uemura, S.

(2006). Context-aware svm for context-dependent in-

formation recommendation. In Proceedings of the 7th

International Conference on Mobile Data Manage-

ment, MDM ’06, pages 109–, Washington, DC, USA.

IEEE Computer Society.

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

ommendation systems. In Brusilovsky, P., Kobsa, A.,

and Nejdl, W., editors, The Adaptive Web, pages 325–

341. Springer-Verlag, Berlin, Heidelberg.

Quercia, D., Lathia, N., Calabrese, F., Di Lorenzo, G., and

Crowcroft, J. (2010). Recommending social events

from mobile phone location data. In Proceedings

of the 2010 IEEE International Conference on Data

Mining, ICDM ’10, pages 971–976, Washington, DC,

USA. IEEE Computer Society.

Ricci, F. (2011). Mobile recommender systems. Informa-

tion Technology & Tourism, 12(3):205–231.

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

302

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

Item-based collaborative ﬁltering recommendation al-

gorithms. In Proceedings of the 10th International

Conference on World Wide Web, WWW ’01, pages

285–295, New York, NY, USA. ACM.

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

(2007). Collaborative ﬁltering recommender systems.

In Brusilovsky, P., Kobsa, A., and Nejdl, W., editors,

The Adaptive Web, pages 291–324. Springer-Verlag,

Berlin, Heidelberg.

Schaller, R., Harvey, M., and Elsweiler, D. (2013). Rec-

sys for distributed events: Investigating the inﬂuence

of recommendations on visitor plans. In Proceedings

of the 36th International ACM SIGIR Conference on

Research and Development in Information Retrieval,

SIGIR ’13, pages 953–956, New York, USA. ACM.

Setten, M., Pokraev, S., and Koolwaaij, J. (2004). Context-

aware recommendations in the mobile tourist appli-

cation compass. In Bra, P. and Nejdl, W., editors,

Adaptive Hypermedia and Adaptive Web-Based Sys-

tems, volume 3137 of Lecture Notes in Computer Sci-

ence, chapter 27, pages 235–244. Springer Berlin Hei-

delberg, Berlin, Heidelberg.

Smyth, B. (2007). Case-based recommendation. In

Brusilovsky, P., Kobsa, A., and Nejdl, W., editors,

The Adaptive Web, pages 342–376. Springer-Verlag,

Berlin, Heidelberg.

Torkington, J. (2014). Small data: Why tinder-like apps

are the way of the future. Retrieved August 13,

2015 from https://medium.com/@janel az/small-

data-why-tinder-like-apps-are-the-way-of-the-future-

1a4d5703b4b.

Woerndl, W., Huebner, J., Bader, R., and Gallego-Vico, D.

(2011). A model for proactivity in mobile, context-

aware recommender systems. In Proceedings of the

Fifth ACM Conference on Recommender Systems,

RecSys ’11, pages 273–276, New York, USA. ACM.

Zhang, W., Wang, J., and Feng, W. (2013). Combining

latent factor model with location features for event-

based group recommendation. In Proceedings of

the 19th ACM SIGKDD International Conference on

Knowledge Discovery and Data Mining, KDD ’13,

pages 910–918, New York, USA. ACM.

Extending Content-Boosted Collaborative Filtering for Context-aware, Mobile Event Recommendations

303