Context-Aware Points of Interest Suggestion with ...

122 downloads 57845 Views 3MB Size Report
POIs, a suggestion about what to do, or an offer, such as a discount given by .... The user / item context is partially specified by the user through the STS Android.
Context-Aware Points of Interest Suggestion with Dynamic Weather Data Management Matthias Braunhofer, Mehdi Elahi, Francesco Ricci, and Thomas Schievenin Faculty of Computer Science Free University of Bozen - Bolzano, Italy [email protected], [email protected], [email protected], [email protected]

Abstract Weather plays an important role in tourists’ decision-making and, for instance, some places or activities must not be even suggested under dangerous weather conditions. In this paper we present a context-aware recommender system, named STS, that computes recommendations suited for the weather conditions at the recommended places of interest (POI) by exploiting a novel model-based context-aware recommendation technique. In a live user study we have compared the performance of the system with a variant that does not exploit weather data when generating recommendations. The results of our experiment have shown that the proposed approach obtains a higher perceived recommendation quality and choice satisfaction. Keywords: Context aware, Weather, Recommender Systems

1 Introduction The decision to purchase a tourism product or to visit a place of interest (POI) is the outcome of a complex decision process. Several factors affect tourist’s decision, some are “internal” to the tourist, such as personal motivators or past experience, others are “external”, e.g., advices or recommendations, information about the product, or the climate of the destination (Swarbrooke & Horner, 2006). In particular, climate and weather are important factors in tourists’ decision making and also influence the successful operation of tourism businesses (Becken, 2010). While tourists might easily predict general climatic conditions, when they travel to a place they will experience the actual weather, which may range among (very) different conditions. For instance, considering the climate factor, it is easy to recognize that Rimini’s beach is much less appealing in winter than in summer, but without knowing the weather conditions a hike on the Dolomites Mountains in summer is hard to miss if the user does not know that a storm is approaching the area. In this paper we focus on systems and techniques that can better predict tourist’s ratings for places of interest, and mobile systems that exploit these predictions in compiling and presenting more appropriate recommendations to tourists. We target this goal by taking into account within the system recommendation model the impact of the weather conditions at a specific place of interest on the (predicted) tourist evaluation for the place. We will show that the knowledge of the weather conditions at a POI, together with past observations of how tourists rated their visited places under several alternative weather conditions may be effectively used to improve the choice satisfaction and perceived recommendation quality of a mobile place of interest recommender system (Ricci, 2011).

This research fits into a fast growing research topic for recommender systems (RS), i.e., context-aware systems (Adomavicius et al., 2011) (Lamsfus et al., 2013). Context-aware recommender systems (CARS) generate more relevant recommendations by adapting them to the specific contextual situation of the user. While the user location and the environment conditions around the user have been often used in (mobile) context-aware recommender system (Ricci, 2011) (Schwinger et al., 2005), the weather conditions that the user will experience at a place of interest have never been systematically exploited in the prediction models of recommender systems. The main difficulty in managing this type of contextual data is that it is not linked to the user state, but depends on the place (item to be recommended) and dynamically varies in time. Moreover, in order to effectively use this type of information the system must be able to predict the weather conditions when the user will visit the place and not simply access these values at the recommendation time. This means that weather predictions data must be obtained and correctly exploited into the prediction model. In this paper we will illustrate the modelling elements of a mobile, context-aware recommender system STS (South Tyrol Suggests) that recommends places of interest in South Tyrol (Italy) by taking into account weather predictions for the places and a matrix factorization model extended with a range of parameters which model the interaction between a set of relevant contextual dimensions and the item ratings (Baltrunas et al., 2012). The tourist’s preference model is learned using a collection of ratings for POIs that the system actively requests to the users before providing recommendations. We have evaluated the proposed system in a live user study by comparing it to a variant (STS-S) that uses exactly the same prediction model and contextual factors with the exclusion of the weather factors. We have found that the exploitation of the weather data can improve the choice satisfaction and perceived decision quality, hence showing the usefulness of the proposed methods.

2 State of the Art Context-aware recommender systems (CARS) is a relatively new field of research but already a number of context-aware rating prediction techniques have been proposed (Adomavicius et al., 2011). Recent research works have focused on model-based techniques that integrate contextualization directly into the rating prediction models, which is the core component of a RS. Applying such techniques, CARS can generate recommendations on the base of a collection of ratings where also the contextual factors, which may have influenced the user experience of the rated item, are specified. For instance, a CARS may record that a user rated five stars a particular attraction (e.g., a museum) after having visited it with his girlfriend in a cold winter day (contextual factors). In this paper we have extended the model originally introduced in (Baltrunas et al., 2011) by incorporating new contextual factors describing the weather conditions and some new user attributes. In fact, while a considerable number of research works have been conducted on CARS, the majority of them have not taken into account one of the most relevant contextual factors, i.e., climate and weather conditions. Several studies have shown that weather condition is a significant factor in tourists’ satisfaction, activity participation, and even in perceived safety at touristic place. Hence, it can strongly

influence tourists’ destination choice (Becken, 2010). Weather condition acts either as a main factor in tourism activities (Kozak et al., 2008) or as a facilitator that makes such activities more pleasant to the tourists (Gómez Martin, 2005). Accordingly, it plays a key role in the national and global tourist flows (Becken, 2010). A study has shown that more than 70% of the German tourists, who have been interviewed, searched and obtained information on the weather conditions of their holiday destinations (Hamilton et al., 2005). Moreover, mobile phones are a main platform for information acquisition and more and more people are using these information and communication devices especially for tourism applications. Device portability makes it easy for tourists to access information in a touristic destination and helps them to find relevant attractions and services, or to support them in the exploration of an area (Ricci, 2011) (Schwinger et al., 2005). Mobile CARS can largely benefit from the exploitation of information relative to the users’ current context. An example of mobile CARS is liveCities presented in (Martin et al., 2011). This system supports tourists by sending them push-based notifications when they enter a certain area and their context matches pre-defined conditions. Such virtual areas with context conditions are created by registered tourism entities and linked with notifications. Such notifications may include the description of nearby POIs, a suggestion about what to do, or an offer, such as a discount given by a nearby restaurant. It is worth nothing that this system is rather difficult to compare with ours. First of all, our system takes into consideration the weather forecast as well as temperature while their system uses only the temperature. Moreover, they use predefined recommendations for the users while we use a predictive model to generate them. Lastly, their system recommends only restaurants while we recommend several types of attractions (events, attractions, accommodation, restaurants and activities),. In (Wang & Xiang, 2012) the authors made a comprehensive analysis of the mobile apps in the travel category of iTunes that have received the largest number of reviews and ratings. They have categorized the 300 top apps according to the unique information services they provide and their design features. Although, several apps provide recommendations using techniques such as collaborative filtering with context awareness, none of them is reported as using weather condition as a contextual factor; this stresses the novelty of our contribution. Finally, in a recent short paper (Meehan et al., 2013) it is presented a work in progress relying on the usage of several contextual factors, including weather information, in a hybrid mobile recommender system that supports tourist’s decision-making process. Even though this system shares some of our goal it was not implemented and therefore it cannot be compare with our system.

3 Usage Scenario This section describes a typical system-user interaction with the implemented places of interest (POIs) context-aware mobile recommender system, named STS. Let us assume that a tourist or a citizen is looking for a POI near to Bolzano. The first time

STS is run it presents to the user the registration screen where she can specify a username, her birthdate and gender. Next, the user is asked to answer the Ten-Item Personality Inventory (10-items TIPI) questionnaire (Gosling et al., 2003) so that the system can assess her Big Five personality characteristics (i.e., conscientiousness, agreeableness, extroversion, emotional stability and openness) (see Figure 1, left and middle). Using the provided birthdate, gender and the assessed personality as input, the system, which implements an active learning component (Elahi et al., 2013), identifies and prompts the user to rate a series of POIs whose ratings are expected to improve the accuracy of the subsequent recommendations (see Figure 1, right). It is worth noting that the requested ratings must be entered by specifying also the context in which the POI was visited; in the considered example the temperature, crowdedness and companion factors should be specified by the user. Now the system can provide recommendations as illustrated in Figure 2 (left). By

Fig. 1. Preference elicitation default, this window provides the tourist with a list of 20 POIs that are considered appropriate for the current context of the recommendation request. We note that some contextual conditions are automatically acquired (e.g., weather conditions), while others can be specified by the user using a system screen (Figure 2, middle). In the event that one of the items is of interest, the user can view its details by clicking on it. In this scenario, the tourist chooses the Dolomites hiking tour as her desired item. Then, the system visualizes a screen with the details of the item, as shown in Figure 2 (right). This window provides various information about the item, such as a photo, its name, a description, its category and more importantly an explanation of the recommendation based on the most influential contextual condition. For instance, in the example “sunny weather” is mentioned as a good motivation for the visit. Other supported features include, among others, the ability to view a POI on the map (Figure 3, left), to write reviews for them (Figure 3, middle) and to bookmark them.

Fig. 2. Context-aware suggestions Bookmarking an item will then allow the user to retrieve the selected POI from the Bookmarked items screen and to be notified when the recommendation for this item become inappropriate due to the weather change at the item’s location (see Figure 3, right). For instance, in our scenario, the next day the weather changes from sunny to rainy. Under this new contextual condition the item is expected to be less suited. Consequently, the tourist is alerted by the system so that she can revise her choice.

Fig. 3. Additional system functionalities

4 Weather-Aware Recommendations STS has been implemented as a rich client, always-on architecture, i.e., the client has been kept as thin as possible and it works only in a limited way offline. The Android client comprises a GUI and a presentation logic component; the entire recommendation logic and data layer reside on the server, which makes use of web services or data storages provided by the Regional Association of South Tyrol’s

Tourism Organizations (LTS1), the Municipality of Bolzano2 and Mondometeo3 in order to obtain the graphical/textual descriptions as well as weather forecast information for a total of 27,000 POIs. The recommendation algorithm of STS computes a rating prediction for all POIs in the database, while assuming that the current or predicted user / item context holds. The user / item context is partially specified by the user through the STS Android client and partially acquired automatically by the system, as it was mentioned in the previous section. More information about the used contextual factors and their contextual conditions can be obtained from Table 1. We have used the contextual factors that were identified in (Baltrunas et al., 2012) where they were selected among a larger number of factors, as being the most relevant and influential ones. Finally, the N highest scoring POIs are included in the recommendation list (20 in our case). In order to generate recommendations in real-time, STS underlying recommendation algorithm follows a model-based approach and has been logically divided into two asynchronous phases: the learning phase and the recommendation phase. The learning phase is performed offline, in our case, once every five minutes. After the model is learned, recommendations can be quickly computed (i.e., in constant time). Table 1. Used context factors Contextual factors and associated contextual conditions Weather Sunny, cloudy, rainy, thunderstorm, clear sky, snowing Temperature Burning, hot, warm, cool, cold, freezing Distance Far away, near by Time available Half day, one day, more than one day Crowdedness Crowded, not crowded, empty Knowledge of surroundings New to area, returning visitor, citizen of the area Season Spring, summer, autumn, winter Budget Price for quality, budget traveller, high spender Daytime Morning, afternoon, night Companion With friends/colleagues, with children, alone, with girlfriend/boyfriend, with family

1 2 3

LTS: http://www.lts.it Municipality of Bolzano: http://www.comune.bolzano.it Mondometeo: http://www.mondometeo.org

Mood Happy, sad, active, lazy Weekday Working day, weekend Travel goal Business, health care, scenic/landscape, hedonistic/fun, religion, visiting friends, education, activity/sport, social event Transport A car, a bicycle, public transport, no transportation means

4.1 Prediction Model The implemented rating prediction model extends and adapts that proposed by (Baltrunas et al., 2012). It is a context-aware matrix factorization approach that incorporates baseline parameters for each contextual condition and item pair, besides the traditional matrix factorization components (i.e., global average, item bias, user bias, and user-item interaction). These baselines are additional model parameters that, for each contextual condition and item pair, measure the deviation of the rating for an item produced by the contextual conditions. For instance, if an item i tends to obtain higher ratings when the contextual condition c holds than when c doesn’t hold, then the baseline for this particular contextual condition and item pair is a positive number. Similarly, if the presence of c leads to lower ratings for i, then the baseline for c and i is a negative number. The introduction of additional parameters, one for each contextual condition and item pair, slightly increases the complexity of the model, however, it improves its accuracy since the model is able to capture the dependency of the ratings from the context (Baltrunas et al., 2011). We have used that approach as a starting point, and we have incorporated additional user attributes (i.e., gender, birth date and personality trait information). Analogously to the extended matrix factorization models described in (Koren et al., 2009), we propose to use a distinct factor vector ya ∈ Rf for each attribute in the set of userassociated attributes A(u) in order to describe the user: 𝑦!

(1)

!∈! !

The matrix factorization model is then as follows: !

𝑏!!! + 𝑞!! 𝑝! +

𝑟!"!!,…,!! = 𝚤 + 𝑏! + !!!

𝑦! ,

(2)

!∈! !

where qi, pu and ya are f dimensional real-valued vectors representing the user u, the item i and the user attribute a, respectively. 𝚤 is the average rating for item i, bu is the baseline parameter for user u, bicj is the baseline for the contextual condition cj and item i, and 𝑟!"!!,…,!! is the predicted rating of user u for item i in the context specified by the conditions c1, .., ck (assuming that k factors are considered). This model,

improving the model proposed in (Baltrunas et al., 2012) and other standard collaborative filtering approaches, allows to produce personalized recommendations based on the aforementioned user attributes, even if the target user hasn’t rated any items yet. 4.2 Training Phase To learn the various model parameters, we minimize the regularized square error on the set of known ratings K (i.e., training set): !

𝑚𝑖𝑛!∗,!∗,!∗,!∗

𝑏!!! − 𝑞!! 𝑝! +

[(𝑟!"!! ,…,!! − 𝚤 − 𝑏! − !,!,!! ,…,!! ∈!

!!!

𝑦! )! !∈! !

(3)

! ! 𝑏!! + 𝑞! !

+𝜆(𝑏!! +

!

+ 𝑝!

!!!

!

+

𝑦!

!

)]

!∈!(!)

Here, λ is the regularization parameter that is used in order to avoid overfitting the observed data. To minimize this equation, we have used standard stochastic gradient descent optimization. 4.3 Recommendation Phase Once the prediction model has been trained, it can generate recommendations by predicting the active user’s rating for each item in the database while taking into account the current or predicted contextual conditions of each item. In order to generate weather-dependent item recommendations, it is necessary to provide the actual or predicted weather and temperature values (for each POI in the data set) as input to the recommendation algorithm. More specifically, the recommendation phase works as follows: 1. The system retrieves the context values for the weather and temperature for the 116 municipalities of South Tyrol by querying the Mondometeo web service4. To improve performance, this information is locally cached on the STS server for 15 minutes. 2. For each item in the database, the system: (a) looks up the item’s location; (b) assigns the weather and temperature values retrieved for the closest municipality to the item; (c) computes a rating prediction, considering the weather and temperature conditions along with other known contextual conditions as input parameters. 3. Finally, the 20 items with the highest predicted ratings are presented to the user as recommendations.

5 Evaluation Recommender systems can be evaluated offline or online (Shani & Gunawardana, 2011). After having fine-tuned the models in an offline stage (not described here for lack of space) we conducted a live user study aimed at measuring the user perceived recommendation quality and choice satisfaction. We compared STS with a simplified 4

Mondometeo web service: http://www.mondometeo.org

version, which does not make use of the weather contextual factors, which is called STS-S. STS-S has a pretty similar interface only lacking the weather related features (weather factors are not mentioned and the prediction model is not using this data). We hypothesized that STS will improve the choice satisfaction and the perceived recommendation quality of STS-S. 5.1 Experimental Methodology In order to prove our research hypothesis, we have designed a specific user task and reused a questionnaire for assessing the perceived recommendation quality and choice satisfaction (Knijnenburg et al., 2012). Moreover, we added a specific question to measure the influence of the weather conditions (see below). 54 subjects (students, colleagues, working partners and sportspeople), aged between 18-35 participated to this experiment, which was conducted in their study or working place. They were randomly divided in two equal groups assigned to STS and STS-S (27 each). They tried and tested the system on a Nexus One mobile phone. We note that, as it is shown below, this sample size was sufficient to prove, with statistical significance, our research hypothesis. The users were invited to imagine that they had an afternoon off, to look for attractions or events in South Tyrol, to consider which contextual conditions were relevant for them and to specify them in the system settings (see Figure 2). Then they were invited to browse the attractions and events sections and check whether they could find something interesting for them. Afterwards, they were invited to browse the system suggestions (recommendations), select the one that they believed fitted their needs and wants and bookmark it. Finally, the users filled up a survey (Knijnenburg et al., 2012), which contains the following statements: Perceived recommendation quality. Q1: I liked the items suggested by the system. Q2: The suggested items fitted my preference. Q3: The suggested items were wellchosen. Q4: The suggested items were relevant. Q5: The system suggested too many bad items. Q6: I didn’t like any of the suggested items. Q7: The items I selected were “the best among the worst”. Choice satisfaction. Q8: I like the item I’ve chosen. Q9: I was excited about my chosen item. Q10: I enjoyed watching my chosen item. Q11: The items I watched were a waste of my time. Q12: The chosen item fit my preference. After this initial interaction, we wanted to assess whether an explicit reference to the weather conditions at the selected item may push the user to change her selected POI. Hence, we offered to the user the opportunity to double check the weather conditions at the selected POI by accessing on a computer the Mondometeo website. This was offered to the users in both groups. Afterwards, we asked the users whether they wanted to change their preferred POI and bookmark another one, i.e., if they believed that, because of the weather conditions at the selected POI, their previous choice was not anymore considered to be appropriate. The users that changed the preferred POI answered the following two questions: Q14: I like the new item I’ve chosen. Q15: Please say again whether the suggested items were well chosen.

5.2 Results The first result of our user study addresses the number of unsatisfied users, i.e., those that have changed their preferred POI. Almost 60% of the users (16 users out of 27) who have used STS-S have changed their selection after the assessment of the weather conditions, while only 30% of the STS users (8 users out of 27) changed their selection. We have obtained p-value equal to 0.028 for the chi square test of independence. This indicates that the usage of the weather conditions in STS has a significant effect on the number of unsatisfied users. In addition, as mentioned before, the users that have changed their selected POI were asked to answer two extra questions, i.e., Q14 and Q15. The result shows that, while the answers of Q14 were almost the same for both systems, there is a significant difference between the answers given to the Q15 by the users in two systems (pvalue=0.01). This means that the users of the STS-S system have expressed significantly lower satisfaction with the second selected item after recognizing that the first item was not appropriate to the weather conditions. Table 2. Perceived Recommendation Quality and Choice Satisfaction Perceived Recommendation Quality

Choice Satisfaction

Q1

Q2

Q3

Q4

Q5

Q6

Q7

Q8

Q9

Q10

Q11

Q12

STS-S mean

3.7

3.4

3.3

3.2

2.7

3.3

2.8

4.3

3.7

3.9

3.5

4.0

STS mean

4.0

3.4

3.5

3.7

2.9

3.8

3.1

4.6

4.0

3.7

3.5

3.9

p-value

0.20

0.56

0.13

0.04

0.14

0.00

0.20

0.02

0.03

0.79

0.42

0.71

Moreover, STS obtained a higher perceived recommendation quality than STS-S. We computed t-test in order to compare the average scores obtained by the two systems (Table 2). We note that statements Q5, Q6, Q7 and Q11 were negatively formulated. Hence, we inverted their scale (in Table 2) for the sake of comparison with the other ones. Among the statements that address the perceived recommendation quality, i.e., Q1-Q7, for two of them we measured a statistically significant different average response in two systems: Q4 (p-value=0.04) and Q6 (p-value=0.001). This means that the users believe that items suggested by STS are considerably more relevant than those suggested by STS-S, and STS users express that they do not like any of the suggested items less than in STS-S. This is a remarkable result and clearly indicates the improvement of STS over STS-S with respect to the perceived recommendation quality. The second part of the questionnaire addresses the users’ choice satisfaction (see again Table 2). Among the statements in this part, two received significantly different answers: Q8 (p-value=0.02) and Q9 (p-value=0.03). Hence, in comparison to STS-S, the users of STS liked significantly more their selected item and were more excited about it.

Finally, we analysed all the survey data in order to discover which system overall scored better with respect to users’ perceived recommendation quality and users’ choice satisfaction. We computed a score for each system by summing up the answers given by each user to each question, and then summing up the values for all the questions. For STS-S we obtained 618 (22.8 per user) total score of users’ perceived recommendation quality and 528 (19.5 per user) total score of users’ choice satisfaction. For STS these numbers were 667 (24.7 per user) and 537 (19.8 per user). Hence, in both cases STS achieved a higher score. We have also compared the total scores of the users of the two systems and got no statistical significance (t-test) for the users’ choice satisfaction (p-value=0.28) but marginal significance for the users’ perceived recommendation quality (p-value=0.07). In conclusion, we can state that the users have perceived an overall higher recommendation quality and choice satisfaction when interacting with STS. Hence, taking into account the weather conditions, our CARS model produces a significant positive effect on the user decision process in comparison to STS-S that does not leverage the weather conditions when generating recommendations.

6 Conclusions In this article we have presented a mobile context-aware recommender system called STS that recommends POIs using a set of contextual factors that include the current and forecasted weather conditions at the recommended POIs. The novelty of this approach rely on the usage of up-to-date weather forecast data into a matrix factorization algorithm to generate personalized context-aware recommendations. We hypothesized that our approach, implemented in STS, improves the choice satisfaction and the perceived recommendation quality of a simpler variant called STS-S that uses exactly the same prediction model and all the contextual factors used in STS except the weather conditions. In a live user study we successfully verified our hypothesis and showed that STS improves significantly the users’ perceived recommendation quality and choice satisfaction of STS-S, hence showing that it is valuable to model and exploit this data into a tourism recommender system. Our future work includes an extended analysis of the data obtained from this study to better understand potential performance differences among the compared CARS, which may be due to the different usage of the weather contextual factors. In addition to that, we would like to test our proposed weather-aware CARS with a larger number of users and a larger rating dataset.

References Adomavicius, G., Mobasher, B., Ricci, F., & Tuzhilin, A. (2011). Context-Aware Recommender Systems. AI Magazine, 32(3), 67-80. Baltrunas, L., Ludwig, B., Peer, S., & Ricci, F. (2012). Context relevance assessment and exploitation in mobile recommender systems. Personal and Ubiquitous Computing, 16(5), 507-526. Baltrunas, L., Ludwig, B., & Ricci, F. (2011) Matrix factorization techniques for context aware recommendation. In Proceedings of the fifth ACM conference on Recommender systems (pp. 301-304). ACM.

Becken, S. (2010). The importance of climate and weather for tourism. Technical Report Lincoln University. Retrieved September, 2013, from http://www.lincoln.ac.nz/PageFiles/6750/WeatherLitReview.pdf Elahi, M., Braunhofer, M., Ricci, F., & Tkalcic, M. (2013). Personality-Based Active Learning for Collaborative Filtering Recommender Systems, AI*IA 2013- XIII Conference of the Italian Association for Artificial Intelligence, Turin (Italy), December 4-6, 2013 Fenton, N. E., & Pfleeger, S. L. (1998). Software Metrics: A Rigorous and Practical Approach, 2nd ed. Boston, MA, USA: PWS Publishing Co., 1998. Gómez Martin, Ma. (2005). Weather, Climate and Tourism. A Geographical Perspective. Annals of Tourism Research, 32(3), 571-591. Gosling, S. D., Rentfrow, P. J., & Swann Jr, W. B. (2003). A very brief measure of the BigFive personality domains. Journal of Research in personality, 37(6), 504-528. Hamilton, J. M., Maddison, D. J., & Tol, R. S. (2005). Effects of climate change on international tourism. Climate Research, 29(3), 245-254. Jamieson, S. (2004). Likert scales: how to (ab) use them. Medical education, 38(12), 12171218. Knijnenburg, B. P., Willemsen, M. C., Gantner, Z., Soncu, H., & Newell, C. (2012). Explaining the user experience of recommender systems. User Modeling and User-Adapted Interaction, 22(4-5), 441-504. Koren, Y., Bell, R., & Volinsky, C. (2009). Matrix factorization techniques for recommender systems. Computer, 42(8), 30-37. Kozak, N., Uysal, M. & Birkan, I. (2008). An analysis of cities based on tourism supply and climatic conditions in Turkey. Tourism Geographies, 10(1), 81-97. Lamsfus, C., Xiang, Z., Alzua-Sorzabal, A., & Martín, D. (2013). Conceptualizing Context in an Intelligent Mobile Environment in Travel and Tourism. In Information and Communication Technologies in Tourism 2013 (pp. 1-11). Springer Berlin Heidelberg. Lewis, J. R. (1995). IBM computer usability satisfaction questionnaires: psychometric evaluation and instructions for use. International Journal of Human-Computer Interaction, 7(1), 57-78. Martin, D., Alzua, A., & Lamsfus, C. (2011). A contextual geofencing mobile tourism service. In Information and communication technologies in tourism 2011 (pp. 191-202). Springer Vienna. Meehan, K., Lunney, T., Curran, K., & McCaughey, A. (2013). Context-Aware Intelligent Recommendation System for Tourism, Work in Progress session at PerCom 2013, San Diego Ricci, F. (2011). Mobile Recommender Systems. Information Technology and Tourism. 12(3): 205-231. Schwinger, W., Grün, C., Pröll, B., Retschitzegger, W., & Schauerhuber, A. (2005). Contextawareness in mobile tourism guides–A comprehensive survey. Rapport Technique. Johannes Kepler University Linz. Shani, G., & Gunawardana, A. (2011). Evaluating recommendation systems. In Recommender systems handbook (pp. 257-297). Springer US. Swarbrooke, J., & Horner, S. (2007). Consumer behaviour in tourism. Routledge. Wang, D., & Xiang, Z. (2012). The new landscape of travel: A comprehensive analysis of Smartphone Apps. In Information and Communication Technologies in Tourism 2012 (pp. 308-319). Springer Vienna.

Suggest Documents