A Context-Aware Model for ProactiveRecommender Systems in the TourismDomain

Matthias BraunhoferFree University of BolzanoPiazza Domenicani 339100 Bolzano, Italymbraunhofer@unibz.it

Béatrice LamcheTechnische Universität MünchenBoltzmannstraße 385748 Garching, Germanylamche@in.tum.de

Francesco RicciFree University of BolzanoPiazza Domenicani 339100 Bolzano, Italyfricci@unibz

Wolfgang WörndlTechnische Universität MünchenBoltzmannstraße 385748 Garching, Germanywoerndl@in.tum.de

Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for third-party components of this work must be honored. For all other uses, contact the Owner/Author.Copyright is held by the owner/author(s).MobileHCI ’15 Adjunct, August 24–27, 2015, Copenhagen, DenmarkACM 978-1-4503-3653-6/15/08.http://dx.doi.org/10.1145/2786567.2794332

AbstractA Proactive Recommender System (PRS) actively pushesrecommendations to users when the current context seemsappropriate. Despite the advantages of PRSs, especiallyin the mobile scenario where users could be provided withrelevant items on-the-fly when needed, the area of PRSsis still unexplored with many challenges. In particular, itis crucial to identify the relevant items for the target usersas well as to determine the right context for pushing theseitems, since otherwise the user acceptance, and thereforesystem success, will be negatively impacted. In this paper,we propose a new model that scores each item on two di-mensions, preference fit and context fit, to proactively pushrelevant items to the target user in the right context. Fur-thermore, we present the preliminary design of a prototypeof a mobile Point of Interest (POI) recommender which willbe implemented in order to evaluate the practicality and ef-fectiveness of our proposed model.

Author KeywordsRecommender Systems; Proactivity; Context-Awareness.

ACM Classification KeywordsH.3.3 [Information Storage and Retrieval]: Information Searchand Retrieval—information filtering

Workshops MobileHCI'15, August 24–27, Copenhagen, Denmark

1070

IntroductionRecommender Systems (RSs) are smart information searchand decision support tools providing users with suggestionsfor items that are likely to be interesting to them or to berelevant to their needs [8]. Proactive RSs (PRSs) are a par-ticular kind of RSs; unlike traditional RSs which deliver rec-ommendations upon an explicit request, i.e., by following arequest-response scheme, these systems make intelligentuse of proactive push notifications to draw the user’s atten-tion to recommendations that are well suited to the currentcontextual situation [10]. For instance, when an art exhibi-tion is opening and the user is estimated to be interestedin art, the system could notify the event to the user whilehaving a break at work.

PRSs offer many advantages in particular for mobile usersin tourism scenarios where the user and Point of Interest(POI) context is highly dynamic (the weather conditionschange, the user continuously moves, the popularity of thePOIs varies with the visiting time), and limitations in screensize and input capabilities make it difficult to browse longlists of recommendations [3]. However, implementing ef-fective PRSs poses a number of challenges. Specifically, inorder for PRSs to gain high user acceptance, and thereforesystem success, they need to ensure that they provide theright information, at the right time, in the right place, in theright way to the right person [2].

In an attempt to address the above challenges, this paperproposes a novel context-aware model for a PRS in thetourism domain. This model predicts and assigns a utilityscore to each POI depending on how well it matches theuser preferences as well as how appropriate is its recom-mendation given the current user and POI context, in orderto be able to push relevant POIs to the target user in theright contextual situation. To this end, it considers various

contextual factors that may affect the user perceived utilityof the pushed recommendation, i.e., the travel time to reachthe POI, the visiting time of the POI, the weather at the POIduring visit, the time available for the user to visit the POI,and the user’s POI visit history.

The rest of the paper is structured as follows: after a briefreview of the related work, we propose a model for proactiv-ity in Context-Aware RSs (CARSs). Then, we describe thepreliminary design of a prototype system that will be imple-mented and evaluated in a user study. Finally, we concludethe paper by discussing the main contributions and futurework.

Related WorkCARSs, which adapt relevant recommendations to the cur-rent contextual situation, are receiving growing attention inrecent years [1]. For instance, [11] presents a CARS thatcomputes recommendations by aggregating the rating pre-diction for each item with a matching score that measuresthe suitability of the item for the current context. A similarCARS based on the same principle is described in [4]. In-terestingly, however, only a few studies have specificallyaddressed PRSs. In one of these studies, [10] proposeda PRS model consisting of two phases. The first phase,called situation assessment, determines whether or not thecurrent contextual situation calls for a recommendation.The second phase, called item assessment, is only exe-cuted when the first phase indicates a promising situationand evaluates the candidate items to finally decide whichitems should be pushed to the user as recommendations.In a subsequent work [9], the authors exploit this two-phasemodel for developing a proactive restaurant RS, and eval-uate the usability of its graphical user interface as well asthe acceptance of its proactive recommendations with testusers. They have found that users consider both widget-

Workshops MobileHCI'15, August 24–27, Copenhagen, Denmark

1071

based and notification-based user interfaces as good so-lutions for implementing proactivity, and that users highlyappreciate proactive recommendations if they are relevantand properly timed.

Unlike the above mentioned two-phase PRS model, whichworks only for one particular type of items (i.e., gas stationsor restaurants) that are similarly affected by different con-textual conditions, our proposed model can handle manytypes of items (e.g., restaurants, mountain hiking routes,museums) that are differently affected by a range of contex-tual conditions. Moreover, our model exploits additional im-portant contextual factors to further enhance the relevanceand utility of proactively delivered recommendations.

Proactivity Model for CARSsBased upon [4] and [11], we propose a model for proactiv-ity that distinguishes between predicted rating and utility,i.e., it assumes that users may like the same item equallyin different contexts but the utility of, in our case, receivinga proactive recommendation for an item differs betweencontexts. For instance, a user usually likes a certain restau-rant or not independently of whether it is close or far away,but the utility of obtaining a proactive recommendation forthe restaurant changes with distance when the user reallyneeds a restaurant. Consequently, the utility depends notonly on the predicted rating (i.e., preference fit) but also onan additional factor, denoted as context fit, that captureshow suitable a proactive recommendation is given the cur-rent user and item context.

To calculate the context fit score for the current context ofuser u and item i, we make use of several heuristics, onefor each considered contextual factor, that assign a valuebetween 0 and 1 depending on how suitable a proactiverecommendation for item i to user u would be when a spe-

cific condition for that factor is observed. We note that theseheuristics can be the same for all users or, if sufficient datais available, can be personalized for individual users orgroups of users. As can be seen in Table 1, our proposedmodel is based on five contextual factors that we conjectureare worth considering when generating proactive recom-mendations for POIs. For example, when evaluating thetravel time from the user’s location lu to the item’s locationli, we propose to use the following heuristic:

ctxF ittravelT ime =

1, lu is identical to lix, x = inverseTravelT ime(lu, li)

0, li is unreachable from lu(1)

where inverseTravelT ime(lu, li) is a negative exponen-tial function of travel time in the range [0, 1] that models theaccessibility of the item’s location [6]. A similar heuristic canbe used to evaluate the pleasure of visiting the item i duringa certain estimated time slot tui (e.g., between 9 a.m. and10 a.m.):

ctxF itvisitingTime =

1, i is most popular during tuix, x = popularity(i, tui)

0, i is never visited/closed during tui(2)

The overall context fit score, denoted as ctxF it, is thenobtained by taking the weighted average of the contextfit scores for the individual contextual factors, where theweights determine the contextual factors’ relative impor-tances. Finally, this overall context fit score can be linearlycombined with the normalized rating prediction (i.e., in the

Workshops MobileHCI'15, August 24–27, Copenhagen, Denmark

1072

Contextual factors Heuristics

Travel time Users do not visit faraway POIsPOI visiting time Users visit certain POIs with

higher probability during certaintime slots

Weather Users demand POIs that are bestsuited for the predicted weatherconditions

User available time Users are looking for POIs theyhave enough time to travel to andvisit

User’s POI visit history Users visit POIs in different pat-terns and at different frequencies

Table 1: Contextual factors used in our proposed model for PRSs

range [0, 1]), which yields the utility of pushing item i touser u:

utility = α ∗ predictedRating + (1− α) ∗ ctxF it (3)

where α ∈ [0, 1] is a meta-parameter that controls the influ-ence of the predicted rating and context fit score on the to-tal utility score. For α values close to 0, the context fit scoreis more important. This is, for instance, useful when the rat-ing prediction is unreliable due to a lack of historical ratingsfor the user or item (cold-start problem). On the contrary,for α values close to 1, the predicted rating gets higher im-portance. The utility values for the items are periodicallyre-calculated, and items whose utility value exceeds a user-specific threshold are candidates to be pushed to the user.

Prototype ImplementationTo evaluate the effectiveness of the proposed approach,we are developing a mobile prototype system that providesusers with proactive recommendations for POIs in the Mu-nich area, as illustrated in Figure 1. It is implemented as arich client always-online architecture, i.e., the client is keptas thin as possible and works only in a limited way offline.The hybrid Android/iOS client leverages Apache Cordova1

and consists of two components: a GUI/presentation logiccomponent, which appropriately notifies the user about in-coming proactive recommendations, and a context-dataacquisition component, which periodically acquires raw con-text data from the device (i.e., location data and calendarentries) and transmits it to the server. The entire recom-mendation logic and data layer reside on the server, whichutilizes third-party web services to retrieve various pieces ofinformation about the POIs (e.g., graphical/textual descrip-tions, weather information, check-ins), as summarized inTable 2. We note that by making use of these web servicestogether with the data that the clients supply we obtain allthe information needed to compute the proposed heuristics.For example, we can infer the visiting time of the POI fromthe expected arrival time and the duration of the visit, wherethe latter can be derived from a statistical analysis of thetimestamps of historical check-ins [7]. Then, it is possibleto determine whether it is proper to visit the POI at this timethrough comparison to the popular POI visiting times thatcan directly be retrieved from the Foursquare API, or, if notavailable, can indirectly be learned from historical check-ins[5]. We are also able to predict how well the POI is suitedfor the weather on the day based upon the weather forecastduring the visiting time, and an analysis of historical check-ins and prevalent weather conditions.

1http://cordova.apache.org/2https://developer.foursquare.com/3https://developers.google.com/maps/documentation/directions/

Workshops MobileHCI'15, August 24–27, Copenhagen, Denmark

1073

Figure 1: Screenshots of the mobile prototype

Discussion and Future WorkIn this paper, we have presented a new model for a proac-tive and context-aware POIs RS to provide mobile userswith relevant POI recommendations at the right time. Thismodel exploits user preferences as well as various heuris-tics to predict the overall utility of pushing a POI recommen-dation to the target user in the current contextual situation.We model the contextual situation with five contextual fac-tors that we conjecture do affect the utility of a proactiverecommendation for a POI, i.e., travel time, visiting time,weather, time available to the user and the POI visit historyof the user.

4https://dev.twitter.com/streaming/overview5https://www.flickr.com/services/api/6http://www.wunderground.com/weather/api

APIs Descriptions

Foursquare2 Gives details about POIs, includ-ing location, opening times andpopular visiting times.

Google Directions3 Provides the estimated distanceand travel time between two lo-cations using a certain means oftransport.

Twitter Streaming4 Allows to filter Foursquare check-ins from Twitter (tweets), whichcan be used to infer the rightweather conditions, visiting timesand visiting durations for differentPOIs.

Flickr5 Allows to retrieve geotagged andtime/date stamped photos of POIs,which can be considered as im-plicit POI checkins.

Weather Underground6 Provides access to currentweather conditions, forecastsand historical weather data.

Table 2: Overview of the used web services

Current and future work includes fine-tuning of the model,finishing our mobile prototype system and conducting a firstuser study, where subjects will be asked to interact withour prototype system and to evaluate the perceived qual-ity of each obtained proactive recommendation in termsof timing and relevance, as well as the usability of the sys-tem. We also plan on using the feedback data collectedfrom the study for designing an improved, more personal-ized model for proactivity, which will be able to better cap-ture differences across users in the perceived satisfaction

Workshops MobileHCI'15, August 24–27, Copenhagen, Denmark

1074

on the timing and content of the proactive recommenda-tions. Finally, we also would like to examine the influenceof additional contextual factors on the quality of proactiverecommendations in the tourism domain, e.g., the healthand fitness level of the user, the companion(s) of the user(social context), or the current activity of the user (such asdriving, cycling and walking).

REFERENCES1. Gediminas Adomavicius, Linas Baltrunas,

Ernesto William de Luca, Tim Hussein, and AlexanderTuzhilin. 2012. 4th Workshop on Context-awareRecommender Systems (CARS 2012). In Proceedingsof the Sixth ACM Conference on RecommenderSystems. 349–350.

2. Gerhard Fischer. 2012. Context-aware systems:the’right’information, at the’right’time, in the’right’place,in the’right’way, to the’right’person. In Proceedings ofthe International Working Conference on AdvancedVisual Interfaces. ACM, 287–294.

3. Damianos Gavalas, Charalampos Konstantopoulos,Konstantinos Mastakas, and Grammati Pantziou. 2014.Mobile recommender systems in tourism. Journal ofNetwork and Computer Applications 39 (2014),319–333.

4. Negar Hariri, Yong Zheng, Bamshad Mobasher, andRobin Burke. 2011. Context-aware recommendationbased on review mining. General Co-Chairs (2011), 27.

5. Hsun-Ping Hsieh, Cheng-Te Li, and Shou-De Lin. 2012.Exploiting large-scale check-in data to recommendtime-sensitive routes. In Proceedings of the ACMSIGKDD International Workshop on Urban Computing.ACM, 55–62.

6. Michael Iacono, Kevin Krizek, and Ahmed MEl-Geneidy. 2008. Access to destinations: how close isclose enough? Estimating accurate distance decayfunctions for multiple modes and different purposes.(2008).

7. Joan Melià-Seguí, Rui Zhang, Eugene Bart, Bob Price,and Oliver Brdiczka. 2012. Activity duration analysis forcontext-aware services using foursquare check-ins. InProceedings of the 2012 international workshop onSelf-aware internet of things. ACM, 13–18.

8. Francesco Ricci, Lior Rokach, and Bracha Shapira.2011. Introduction to recommender systems handbook.In Recommender systems handbook. Springer, 1–35.

9. Daniel Gallego Vico, Wolfgang Woerndl, and RolandBader. 2011. A study on proactive delivery ofrestaurant recommendations for android smartphones.In ACM RecSys Workshop on Personalization in MobileApplications, Chicago, USA.

10. Wolfgang Woerndl, Johannes Huebner, Roland Bader,and Daniel Gallego-Vico. 2011. A model for proactivityin mobile, context-aware recommender systems. InProceedings of the fifth ACM conference onRecommender systems. ACM, 273–276.

11. Wolfgang Woerndl, Henrik Muehe, Stefan Rothlehner,and Korbinian Moegele. 2010. Context-awarerecommendations in decentralized, item-basedcollaborative filtering on mobile devices. In MobileComputing, Applications, and Services. Springer,383–392.

Workshops MobileHCI'15, August 24–27, Copenhagen, Denmark

1075