Context-aware User Interaction for Mobile Recommender Systems

Abstract

We are witnessing more and more mobile recommender systems applied in e-tourism, utilizing contextual information in order to boost productivity and user experience by generating more tailored suggestions to them. In the meantime, the design of these applications is getting more and more user-centered. Context of use is playing an important role in the appropriateness of a user interface (UI). This motivates our study to devise a novel approach for augmenting user interaction experience on smartphones by exploiting the current context of the end user. For this purpose, we have designed and implemented a context-aware UI based on an existing mobile application for recommending tourist places. We have conducted a user study and measured the e ectiveness of our method in terms of three attributes: task completion time, the perceived user ease of use and the perceived user satisfaction. The within-subject evaluation conducted with 25 participants con rms that the proposed contextaware UI can enrich user interaction experiences. User interfaces; mobile application; recommender system; tourist guide; user study

1. INTRODUCTION

For many tourists, smartphones and other mobile devices have become indispensable as electronic personal tour guides, since they enable the access to tourist information and services anytime and anywhere. Particular emphasis has been given to bene ts of exploiting contextual information in applications to provide highly accurate and relevant services to users. However, utilizing these contextual parameters to update the user interface (UI) based on the user's current contextual situation has not been given much consideration in previous studies. The fundamental motivation for our work is to investigate the in uence of exploiting the traveler context in which the mobile tourism application is intended to be used in order to o er a situation-aware UI.

The primary hypothesis of the study is whether tailoring the application's UI and interaction methods to the current context of the end user enriches the end user's interaction experience. Because by considering context of use, more supportive input modalities might be o ered to users to assist them for their current situation. To prove this hypothesis, in this paper, we propose and test di erent UI options in certain contextual situations such as moving context and still context. We also present the results of a conducted user study in which users were asked to test the adapted UI and compare it with the original, non-adapted UI. In this user study, we logged the user interaction and recorded the task completion times. Besides that, we measured the perceived user satisfaction and ease of use by means of an online questionnaire.

The contents of this article can be summarized as follows: Section 2 discusses related work and focuses on the existing applications. We will then describe the solution and highlevel design and elaborate the technical set-up of the experiments in Section 4. We designed and implemented our work upon of a fully functional application that is called South Tyrol Suggests (STS) [ 5 ], which we will describe in more details in Section 3. Section 5 is dedicated to evaluation results of our experiment and a discussion of the proposed novel methodology addressing the research concerns. Finally, Section 6 draws conclusions and provides future work directions. 2.

RELATED WORK

In this section, we investigate the current state of the art of: 1) the available interaction methods in existing applications, 2) and the literature and other approaches proposed for tourism recommender systems (RSs) on smartphones. 2.1

Interaction Methods on Existing Apps

Due to the lack of more exible and personalized gestures for interaction with smartphones and the limitation of their screen sizes and other resources to access rich functionalities, new methods of input for mobile devices should be investigated and expanded. To resolve the aforementioned issue, extensive research has been conducted and a large number of interaction modalities has been proposed. Brewster et al. [ 6 ] studied pressure input method in their work. They designed a pressure-based keyboard where a soft press on the touchscreen generated a lowercase letter and a harder press an uppercase one. The results showed that text entry with pressure keyboard was e ective and can outperform a standard shift-key keyboard design on a mobile touchscreen device.

In addition, as nowadays most of the tasks are being done mobile, we need more interactions that leave visual attention unoccupied, so that users can concentrate on their main activity. One of these eyes-free input modalities is auditory interfaces presented by Sodnik et al. [ 19 ], which is useful for non-concentrated situations like driving. The proposed approach proved to be very e ective and safe to use, since the driver could have a low-level distraction compared to the traditional user interactions. Likewise, in our previous work [ 13 ], we investigated three di erent input methods in such non-concentrated scenarios. We compared three different input methods and we demonstrated free-form gestures such as tilt in our application outperformed other input modalities when the environment is distracting, and are more embraced by users in spite of higher noise rate to the cause of non- concentrated situations' nature.

Authors in [ 14 ] expanded the bandwidth of new interaction methods by proposing rotation of a handheld device around a single axis in a 90 degree range to make choice among menu items. The system is immediately learnable by novices and supports eyes-free use by experts.

Previously, Dybdal et al. [ 9 ] investigated hands-free interactions using eye movements only to control mobile devices. Two gaze-based strategies, i.e., dwell time selections and gaze gestures, were compared in their experiment to discover the optimal interaction with smartphones. The main advantage of gaze interaction is that it can be done handsfree, thus allowing to control a mobile device without touching it. They demonstrated, gaze gestures are less error prone and were faster than dwell selections by gaze but generally gaze interaction had a lower performance than touch interaction.

Another interesting technique proposed to enhance the performance of user interaction on smartphones is voice. Sakamoto et al. [ 18 ] suggested a technique called voice augmented manipulation (VAM) to augment a user's input (a nger gesture or button press) with voice input in a mobile device. There were two methods: one used a user's voice and nger gestures (scrolling, pinching, etc.), and the other used a user's voice and a button interface.

Recently, Rozado et al. [ 17 ] also carried out a study to test the feasibility of using gaze gestures to interact with a smartphone. They compared two modalities of performing the gesture: without dwell-time and with dwell-time. Based on a data set collected from 20 participants, the modality without using dwell had faster completion time, yet less accurate and more error probable than the modality using dwell time. Hence, learning e ects have been studied in this project, which revealed no obvious learning e ect over time in accuracy or performance.

Chen et al. [ 7 ] presented BlindPass, a new password entry method. They developed and evaluated four eyes-free password entry methods: (i) Number Pad, which augments the conventional input methods such as number pad with auditory feedback, tactile feedback, or both; (ii) Wheel, which mimics the spatial location of numbers in the clock face. Users can tap the numbers according to their spatial difference, or swipe along the screen to enter a password; (iii) Stroke, which allows users to enter the password using gesture, similar to a marking menu; and nally (iv) Scroll, where users input the password by swiping up or down from any position on the screen until the desired number is reached before lifting up their nger; on a smartphone in order to enhance security in secured applications such as e-banking, booking ight tickets, and etc. Their results showed that eyes-free passwords are easy to use and relatively easy to learn. Further, both Wheel and Stroke input methods were faster to perform.

Prior studies have explored a variety of input modalities, however, most of the work discussed so far is not for tourism and/or mobile domain. We focused on applying some of the user-centered, exible gestures which could possibly improve tourists' mobile interaction satisfaction. 2.2

Tourist Recommender Systems on Smartphones

The capabilities of modern smartphones make them a primary and indispensable platform in people's everyday life. One of their most attractive services that people take advantage is mobile tourist guide applications wherein tourists increasingly spend considerable time planning their travel activities. Therefore, we review some of the research in this area that in uenced our solution. New developments in network connectivity and the wide variety of available sensors in current mobile devices gave rise to the eld of contextaware application and services, assisting applications to o er more personalized and tailored services to tourists in order to address some of the limitations of handheld devices such as small size and limited processing power.

Lately, Gavalas et al. [ 10 ] proposed a classi cation of existing mobile tourism RSs prototypes on the basis of the following aspects: (a) their chosen architecture, (b) the degree of user involvement in the delivery of recommendations, and (c) the criteria taken into account for deriving recommendations. Furthermore, they investigated the last item in more detail and classi ed as following: user constraints-based recommender systems, pure location-aware recommender systems, context-aware recommender systems and critique-based recommender systems. At the end the potential trends for this eld in addition to its challenges have been provided.

One of the earliest works that applied context-awareness in the mobile tourist application was Setten et al. [ 21 ] in 2004. They combined context-awareness with recommender systems in a mobile tourist application named COMPASS which serves a tourist with map-based information services based on their interest and some speci c context.

Another context-aware mobile recommender project, i.e., ReRex [ 2 ], has investigated the importance of exploiting a traveler's contextual situation for recommending pointsof-interests (POI) on mobile recommender systems and assessed user acceptance of these recommendations by asking users to judge whether that contextual factor actually affects their rating. They took into account several important contextual factors which seem more e ective in the generation of the relevant and personalized recommendation and shortly justi ed the recommendations. The results revealed the higher user acceptance and satisfaction for the contextaware version.

In 2013, Braunhofer et al. [ 5 ] initiated the development of South Tyrol Suggests (STS), a context-aware mobile recommender system that suggests POIs in South Tyrol in Italy. STS provides several innovative interface elements, including: (a) personality questionnaire, i.e., a brief and entertaining questionnaire used by the system to learn the user's personality; (b) active learning module that acquires context-dependent ratings for POIs that users are likely to have experienced, hence, reducing the stress and annoyance to rate (or skip rating) items that the users do not know [ 4 ]; and (c) recommendation module that relying on matrix factorization leverages context-dependent ratings and personality information in order to generate more relevant personalized recommendations for users, even if they are new to the system (i.e., new user cold-start problem). In several live user studies and analyzing the log data produced by a larger sample of users that have freely downloaded and tried STS through Google Play Store, the authors have evaluated the system and showed that the system in general is perceived as useful and easy to use.

Biuk-Aghai et al. [ 3 ] designed and implemented contentbased recommender systems for tourists. They extended their previous app "MacauMap" (a mobile tourist guide and map system) and employed a genetic algorithm for generating travel plan and a fuzzy-logic based module for calculating visit/stay times for each stop of the entire trip.

In order to o er more customized items to travelers, Kularbphettong et al. [ 12 ] introduced hybrid recommendation on smartphones including ontology, collaborative- ltering and location-based services methodologies. They designed and implemented a heritage-tourism mobile recommender system and assessed the suitability of places for users by using aforementioned methodologies.

Tumas et al. [ 20 ] implemented a personalized mobile city transport advisory system (PECITAS). Using this app users can obtain recommendations for personalised paths between two arbitrary points in the city of Bolzano, Italy by city transport means and walking which generates multiple routes. They used knowledge-based technology to recommend and rank di erent routes more personalized to the guests by exploiting their travel-related preferences.

Pawara et al. [ 16 ] implemented and tested a prototype which is an example of systems using collaborative- ltering. The application took advantages of collaborative user-generated content, which has a location-aware chat system. They claim that their application has easier access to information than those that used social network as it does not require process of joining and requesting associations so as to approach the valuable information of other tourists.

Page et al. [ 15 ] discovered that a communication style personality trait they called "for your information" (FYI) mostly predicts the adoption of location-sharing social networks (LSSN) and disclosure behavior. LSSN enable users to share their location with their family and friends and bene t of social advantages. They also found that the youngest interviewees are commonly FYI communicators.

One of the recent works which utilizes contextual information for recommendation is Pythia, a privacy-enhanced personalized contextual suggestion system for tourism. An innovative user-centric architecture has been proposed in this work which combines the following features: the (sensitive) personal data (e.g., location data) are stored at the user-side as well as the pro le of user interest which will be created based on these data. The contextual suggestions are also generated at the user-side. This combination o ers strong privacy since the personal data is not disclosed to any party, including the recommender service provider [ 8 ].

The state of the art stated in this section reveals that mobile tourism RSs are one of the most popular strand of research for mobile RSs. Particular emphasis has been given to bene ts of exploiting contextual information to deliver highly accurate and relevant recommendations. Despite these signi cant e orts, not much work has been done with respect to exploit these contextual parameters to update the UI based on user current situations. This highlights our study motivation that takes into account traveler current context to o er situation-aware UI. 3.

APPLICATION SCENARIO

We designed and implemented our work upon South Tyrol Suggests (STS) [ 5 ]. STS is an Android-based mobile recommender system aimed at o ering context-aware recommendations for touristic items (i.e., accommodations, restaurants, sport, cultural attractions and events) for the South Tyrol region of Italy. Its main functionality is to suggest context-aware and tailored recommendations of touristic items to tourists and they can search for POIs.

This application uses a repository of approximately 27,000 POIs/tourist items data, as it is connected to the most comprehensive databases o ered by the Regional Association of South Tyrol's Tourism Organizations (LTS1), the Autonomous Province of Bolzano2, the Municipality of Bolzano3 and SASA4.

The system computes rating predictions for items by considering their contextual information with the aim of providing more accurate and personalized recommendations for tourists. Moreover, user personality and active learning are exploited in order to tackle the cold-start problem. Then, the items with the highest predicted ratings are recommended for that speci c context.

As depicted in Figure 1, the rst phase of the active learning procedure in STS starts by entering some basic information by the user through registration stage such as her birthdate and gender, in addition to the Five-Item Personality Inventory (FIPI) [ 11 ].

Another source of information to predict the most relevant items for the user can be obtained from the context settings, which are accessible from the user pro le page, as illustrated in 2. They allow the user to ne-tune the current contextual situation by enabling and setting the values of those factors that can not be automatically acquired, such as the duration of the current stay, the user knowledge of the travel area, the current budget, the actual companion and feelings.

By exploiting the evaluated personality (as mentioned previously), the user's age and gender (if available), in addition to the value of the considered contextual factors, the system 1LTS: www.lts.it 2Autonomous Province of Bolzano: www.provinz.bz.it 3Municipality of Bolzano: www.gemeinde.bozen.it 4SASA: www.sasa.bz.it identi es and shows 20 highly relevant POIs to the tourist. The user can see for each item on the list information such as a photo, its name, as well as an explanation of the reason why that recommendation has been o ered which estimated most in uential contextual condition by the system. Another feature of the POI suggestions screen is that it provides users with a pop-up window that request the user to provide more ratings-in-context for POIs, as can be seen in Figure 3.

As illustrated in Figure 4 (left), by clicking on any preferred item, the user is redirected to the item details page where she can access more data of the selected POI such as its photo, name, description, user reviews, its category as well as an explanation of why this item was o ered to the user or request a route suggestion to reach there. Tagging and bookmarking the POI in order to easy get back to it later is also possible in this page. Furthermore, there are options to rate the item as well as write a review for it, which is depicted in the right image.

Our work enriches the app by considering context such as tourists' current activity like walking or still in order to tailor the UI. 4.

DESIGN AND IMPLEMENTATION

This section provides a description of the research approach used in this project in addition to other possible solutions that might be useful. At rst, the initial ideas are explored, and then we elaborate on the technical setup of the described experiments in this section. 4.1

Research Methodology

The main idea is tailoring and adopting input modalities to interact with handheld devices based on tourist's current contexts. In our previous work [ 13 ], we investigated the inuence of interaction methods on the user's rating behavior as one possible source of noise in ratings. In the next stage, we would like to address the following issues:

Automatic context detection

Introduce new interaction methods

Set an appropriate modality for the detected context Afterwards, we implemented the solution into STS for the experimentation. Finally, the perceived user satisfaction was evaluated to measure the usefulness of our proposed approach. 4.2

Contextualizing Input Methods

It is important to identify meaningful contextual factors, e.g., ambient conditions, display brightness level, display size, noise level, user activity, experience of user (experienced or new). Also it is crucial to identify possible interaction consequences (e.g., interface will change from colour to black and white, font or picture size increases, certain functionality like voice input / output is omitted, certain functionality asks for con rmation (e.g., "5 stars? Are you sure?"). Users could then be asked to perform a speci c user task under various contextual conditions (can be simulated) using a speci c interaction method (e.g., search for a POI recommendation, provide a review for a POI) and provide their subjective feedback, which together with some implicit feedback (time to completion, success rate, ...) will be used to evaluate the right interaction method for each context.

Table 1 has been obtained by analyzing some other scienti c methods in addition to suggesting some new context and interaction consequence(s) pairs to examine. Zheng et al. [ 22 ] have proposed a rule-based approach in order to design the context-sensitive mobile UI. The rules consisted of conditions and actions. In order to personalize their proposed context-sensitive UI, they have considered the context information that comes from various sensors built in the mobile device as well as from the user's pro le. Suppose, for instance, a user wants to check the latest recommendations and order an item in a dark and noisy environment. Then, the UI can display the result in large font text, increase the brightness level of the screen; accept the order given by the user by tapping, rather than typed on the keyboard, or by voice in order to adapt to the context. In this scenario, the rule was: change to large text output while the environment is dark and noisy. Besides, there were three options to accept the order given by the user including tapping, typing on the keyboard, or by voice. Each action would be activated based on the match context value. Therefore, tapping would be activated in her current situation.

Our research di ers in that we have been able to implement a context-adaptive UI and conduct a user study to investigate whether it enriches the end user interaction experience on smartphones.

In the following, we are describing the suggested context and UI consequence(s) pairs:

To improve the accuracy of the user rating we propose to ask her for con rmation after entering a rating, while she is stressed by high noise levels. For example, "1 star? Are you sure?"

On-screen tutorials/context-sensitive help can quickly show new users what's important on the Application in use. Hence, gestures used for the interaction and the UI controls can also be explained.

For some situations like driving it is very important employing interactions to not draw the user's visual attention away from their main activity. Sodnik et al. [ 19 ] proposed exploiting auditory interfaces, which are eyes-free gestures and proved they have a low-level distraction compare to the traditional user interactions. Therefore, these interfaces are very e ective and safe to use.

When a user is in distracting situations such as walking, it is hard to focus on the mobile screen. Thus, bigger icons/buttons and large font size was proposed to keep the user focused on her main activity. In contrast, in a sitting situation she can further concentrate on her mobile screen so representing more details might be preferred.

In high noise level conditions, detecting the correct command via voice is very challenging. So, omitting the voice gesture is recommended in these situations.

Another case is when the weather is cold; touchscreen gestures would be very annoying. Since users do not like to be enforced taking their gloves o and freezing their ngers o to use the smartphones. Hands-free gestures like voice or gaze have no such problem.

Rozado et al. [ 17 ] indicated the bene ts of using gaze gestures in situations where one or both of our hands are busy with other tasks. For example, when using public transportation and holding the smartphone with one hand while using the other hand to hold a handle, it would be very useful to send commands to the phone by means of gaze gestures.

In addition, downloading videos are only suggested if the mobile phone is using Wi-Fi signals. In other Internet connections (e.g., 3G) the application only displays images [ 22 ].

Meanwhile, in our previous study [ 13 ], we illustrated that di erent users might prefer di erent interaction methods in various contexts. As a result, a dynamic adaptation of the UI based on individual users is preferred. The appropriate UI for the detected context and the user type must be learned by the system, not simply enforced by the system designer. Thus, there is the need to implement learning techniques that detect what input modality or UI element quali es that speci c situation as well as the user type. This could be done as a classi cation task where its input comprises the user personality type and the corresponding context, and the category/class is one of the available input modalities/UI features.

We are in the phase of proposing and testing the several options in the certain contextual situations, which can be the foundation of a dynamic adaptation. In other words, this information can be utilized as example inputs for personalization.

We created Table 1 as a starting point with options in principle. However, we opted to implement only some of them which will be represented in the Table 2. A user study will allow to determine which interaction consequence(s) are truly relevant for the various contexts.

4.3 Implementation Details

Table 2 speci es the augmented functionalities for the STS app in this study.

One of the most important contexts for the study purpose is the user's current activity. To this objective, we used Google's Android Activity Recognition API [ 1 ] to recognize a user's current activity, such as walking, driving, or standing still. 4.4

User Interface

The screen where users are asked to bookmark an item is illustrated in Figure 6. We proposed long press anywhere on the item detail page as a bookmark interaction method in moving context which is shown in Figure 6b. Figure 6a shows how to bookmark an item in original UI before applying context-awareness.

As shown in Figure 7 there are three options for rating. For moving context, we suggested removing pop-up review page and tilt gesture for rating in the main item detail page (Figure 7c). In addition, one- nger-hold-pinch was proposed for sitting context on the pop-up review page (Figure 7b). This gesture is a two- nger gesture. One nger is kept on the screen, while the second nger goes farther or closer on the screen to increase or decrease the rating stars [ 13 ]. Figure 7a also shows the original UI for rating which works by touching the star on the review pop-up page. 5.

USER STUDY AND EVALUATION RESULTS

To evaluate the usability and e ectiveness of our proposed method, we did an experiment to examine how the quality of user interaction in a mobile recommender system environment is in uenced by the context-adaptive, multimodal UI based on the user's current context. Another objective of this study was to study the learning curve, i.e., to see whether task completion time or other metrics are lower/better (changes) after n iterations (trials) with the same interaction method. At the end we evaluated the usefulness of our proposed approach and the e ciency of the opted interaction method for each context condition by assessing the objective criteria (e.g., task completion time) as well as subjective measures (i.e., perceived user satisfaction etc.). The tests that were performed include a main comparison: STS original UI (non-context-aware UI) versus context-aware UI.

Side hypotheses of the research are de ned as follows: H1 Bigger icons in moving context reduce the task completion time; H2 Bigger icons in moving context have a direct positive e ect on perceived user satisfaction; H3 Long press anywhere on the item details page in order to bookmark that item in moving context reduces the task completion time; H4 Long press anywhere on the item details page in order to bookmark that item in moving context has a direct positive e ect on perceived user satisfaction; H5 The review question page before rating is omitted and switch to tilt interaction to rate in moving context reduces the task completion time; H6 The review question page before rating is omitted and switch to tilt interaction to rate in moving context has a direct positive e ect on perceived user satisfaction; H7 One- nger-hold-pinch gesture for rating function in still context reduces the task completion time; H8 One- nger-hold-pinch gesture for rating function in still context has a direct positive e ect on perceived user satisfaction; H9 There is a learning curve in every interaction method; especially in complex input modalities the learning curve has a higher value, e.g., task completion time or other metrics are lower/better after n iterations (trials) with the same interaction method; (a) (b) (c)

The study consisted of one experiment with two conditions. One condition was performed with the proposed contextualized UI version, whereas the other condition was performed with the original STS before applying context-awareness to the UI.

Participants were rst introduced to the application and scenario in addition to the determined tasks to accomplish. Afterwards, respondents were asked to search among suggested items while walking. Since the detected context by the application is walking, the UI switched to bigger icon/ buttons. Then they were asked to select an item, bookmark it and rate the item. Long press on the screen is applied to bookmark an item. As regards to the rating, the review question page before the rating page is omitted and the input method switched to tilt gesture. All rules for walking context have been applied in this step for all aforementioned functions. Then they were asked to rate items in still context while one- nger-hold-pinch is activated. In addition, we asked participants to do aforesaid steps 10 times in order to measure the learning e ects while performing the tasks. Their task completion time has been logged for each function separately.

In order to evaluate our hypothesis, we used the same study design and questionnaire with both app versions. To avoid bias due to possible learning e ects, we used a counterbalanced experiment. One group rst tested the contextaware UI, then the original UI; another group did it viceversa.

After performing all determined tasks, users were asked to ll out an online questionnaire. The questionnaire aimed at evaluating the proposed context-aware UI from the end user point of view as well as how users perform on realistic tasks. The questionnaire consisted of four main categories: prior knowledge, context-adaptive UI perceived user satisfaction, original UI before applying context-awareness perceived user satisfaction, improvement of interaction experience from end user point of view. In each part we inquired the ease of use as well as the user satisfaction for each method. At the end the interviewer asked participants if the user interface and interaction methods tailored to their current context improved their interaction experience, and if they have further suggestions or ideas on the proposed idea. 5.2

Participants and Apparatus

We conducted a within-subjects user study in order to test the hypotheses. The experiment followed a within-subjects design which means every single participant is subjected to every experimental treatment, thus allowing also a small sample of respondents. It involved 25 participants (selected from the computer science and mathematics student population at the Technical University of Munich) aged between 2235. The experiment was performed using a Samsung Galaxy S6 mini smartphone running Android 5.1. 5.3

Evaluation Results

As mentioned in the procedure of the experiment in Section 5.1, after the users tested the UIs, they answered the set of questions related to the perceived user ease of use and satisfaction. The questions were evaluated using a Likert scale with ve options ranging from "Very dissatis ed" with the value of 1 to "Very satis ed" with the value of 5 (0 means the user did not rate that UI). 5.3.1

Perceived Ease of Use

Figure 8 provides information about how satis ed users were with the ease of use of the four mentioned functions in the context-aware and original UIs. The data represented are the mean value of the user responses for the ease of use, in addition to the error bars, which indicate the standard deviation of these data.

As can be seen from the data, the context-aware UI for searching among the items and rating the items in moving context performs noticeably better than the original UI. There is slightly less di erence of perceived user ease of use for rating items in sitting context between context-aware UI and original UI. For the item bookmarking function in moving context, the pattern is repeated. Furthermore, the bookmarking function of the original UI has the highest value of standard deviation.

In general, users are more satis ed with the ease of use in the context-aware UI compared with the original UI in the determined functions in moving and sitting context. The Search in Moving

Rate in Moving Bookmark in

Moving

Rate in Sitting Context-aware UI

Original UI one-tailed t-test shows that the di erence in ease of use between the context-aware UI and the original UI is statistically signi cant (p-value = 0.05). The di erence in rate in sitting is marginally statistically signi cant (p-value = 0.10). Rate in moving is very statistically signi cant (pvalue = 0.003). Besides, bookmark in moving is also likely to become statistically signi cant if the sample size is increased. 5.3.2

Perceived Satisfaction

Figure 9 shows the mean value of perceived user satisfaction for the di erent functions in both context-aware and original (non-context-aware) UIs. The questions are evaluated using a Likert scale with ve options ranging from "Strongly disagree" to "Strongly agree". Analogously to the previous chart, the error bars indicate the standard deviation of these data. 6   n o it fca 5   s it a rS 4   e s U ed 3   v i e c erP 2   f o e lau 1   v n ea 0   M context-aware UI can lead to a better interaction between users and smartphones because interaction methods are tailored to the users' current context.

Regarding the bar chart, the highest user satisfaction was obtained by the function bookmarking the items in the contextaware UI, closely followed by searching among items in the same UI. These results are in accordance with hypotheses H4 and H2. By contrast, satisfaction result for searching among items in the original UI is considerably lower. Also, perceived user satisfaction of rating items in the contextaware UI in both moving and sitting contexts is higher than in the original UI which indicates hypotheses H6 and H8.

To sum up, users have demonstrated higher satisfaction of the context-aware UI in comparison with the original UI. Based on one-tailed t-tests for the perceived satisfaction of the context-aware UI and the original UI for both search and bookmark in moving context, the results are considered to be statistically signi cant (p-value = 0.01) for both of them. On the other hand, rate in sitting and moving context are marginally statistically signi cant. It might be remedied by testing among more participants. 5.3.3

Task Completion Time

The graphs in Figure 10 show in seconds (s) terms the changing patterns of task completion time for the aforementioned functions for both context-aware and original UIs over the 10 measurements.

As can be seen in Figures 10a and 10b, the task completion time for the context-aware UI is lower than for the original UI which supports hypotheses H1 and H3, respectively. By contrast, in Figure 10d, the pattern is reversed which is in contradiction with the hypothesis H7. We think improving the implementation or altering the selected input modality might solve this. Furthermore, Figure 10c shows higher task completion time for context-aware UI in rst measurement whereas it is equal or lower for the next measurements similar to the hypothesis H5.

It is interesting to note that regarding the four graphs, there is one basic general trend over the measurements: downward and then leveled o . It shows that there is a learning curve in every interaction method; specially in complex input modalities the learning curve has higher value, i.e., mean task completion times are lower after n iterations (trials) with the same interaction method which is according to our hypothesis H9. 6.

CONCLUSION AND FUTURE WORK

In this study we built a context-aware UI to help users to boost their interaction experience for the South Tyrol Suggests (STS) mobile tourism RS. We have investigated how adapting the UI to the tourist context of use can improve her satisfaction. To this end, we used the Google's Android Activity Recognition API to recognize the user's current activity whether she is in still or moving context. We developed an appropriate UI depending on that speci c context. Users then were asked to perform tasks in a study under those contextual conditions and provide their subjective feedback (i.e., perceived user satisfaction), which together with some implicit measures (task completion time) have been used to evaluate the research hypothesis.

We showed that the context-aware UI outperforms the original UI in terms of both perceived user ease of use and perceived user satisfaction. Besides, the statistical signi Search in Moving

Rate in Moving Bookmark in

Moving

Rate in Sitting Context-aware UI

Original UI

We conjecture that the context-aware UI has a direct positive e ect on perceived user satisfaction compared with the original UI. We are interested in this hypothesis as we think RateinSitting 0  1  2  3  4  5  6  7  8  9  10 

Numberofmeasurements (d) 7  6  )5  i(se m tn4  iltepom3  scok aT2  1  0  1  2  3  4  5  6  7  8  9  10 

Numberofmeasurements (a)

RateinMoving 8  7  tit()se6  im5  on le4  scaopT2  m k3  1  0  1  2  3  4 Numb5e rofm6e asure7m ents8  9  10  (b)

BookmarkinMoving 4  3,5  it)(s 3  em2,5  n tieo 2  l p com1,5  sak T 1  0,5  0  1  2  3  4  5  6  7  8  9  10 

Numberofmeasurements (c) OriginalUI Context-awareUI OriginalUI Context-awareUI OriginalUI Context-awareUI OriginalUI Context-awareUI cant test showed the outperformance for the rating function in walking state in terms of the perceived ease of use. In addition, searching and bookmarking functions in moving context in terms of the perceived satisfaction showed significant di erences. We conjecture the perceived satisfaction may be improved by involving further participants.

In addition, the context-aware UI decreased the task completion time in search, rate and bookmark functions in moving context. We believe, a more sophisticated implementation of both interaction methods and tourist context detection could help to further reduce task completion time. Another possible extension might be implementing and evaluating other input modalities for each aforementioned functions to decrease task completion time.

It is noticeable that regarding the task completion time graphs, a learning curve was observed during the experiment. This means after repeating the same task in a series of trials the learning will be increased; consequently task completion time can be reduced.

We can conclude that the proposed context-adaptive UI outperformed the original UI of STS that did not utilize the current context in which the application is intended to be used. Since context of use is playing a critical role in the appropriateness of a UI, further investigation of this approach is strongly recommended.

Future work can be done in a number of di erent directions. In a practical sense, the current quality of our context detection should be improved and some more contexts such as light, noise, weather and Internet connection of the environment should be considered.

Future research could include improvements to the system personalization, i.e., UI adaptation to context should not be same for all users. So the system must learn the type of user, in terms of adaptation preferences. In other words, as we discussed di erent users might prefer di erent interaction methods in di erent contexts. Therefore, our future work lies on using learning techniques for personalization. This could be done as a classi cation task where its input comprises the user personality type and the corresponding context, and the category/class is one of the available input modalities.

References

[1] https: //developers.google.com/android/reference/com/ google/android/gms/location/ActivityRecognitionApi.

[2] L. Baltrunas , B. Ludwig , S. Peer , and F. Ricci . Context-aware places of interest recommendations for mobile users . In Design, User Experience, and Usability. Theory, Methods, Tools and Practice , pages 531 { 540 . Springer, 2011 .

[3] R. P. Biuk-Aghai , S. Fong , and Y.-W. Si . Design of a recommender system for mobile tourism multimedia selection . In Internet Multimedia Services Architecture and Applications , 2008 . IMSAA 2008 . 2nd International Conference on, pages 1 {6 . IEEE, 2008 .

[4] M. Braunhofer , M. Elahi , M. Ge , and F. Ricci . Context dependent preference acquisition with personality-based active learning in mobile recommender systems . In Learning and Collaboration Technologies. Technology-Rich Environments for Learning and Collaboration , pages 105 { 116 . Springer, 2014 .

[5] M. Braunhofer , M. Elahi , and F. Ricci . Techniques for cold-starting context-aware mobile recommender systems for tourism . Intelligenza Arti ciale , 8 ( 2 ): 129 { 143 , 2014 .

[6] S. A. Brewster and M. Hughes . Pressure-based text entry for mobile devices . In Proceedings of the 11th International Conference on Human-Computer Interaction with Mobile Devices and Services. ACM , 2009 .

[7] C. Chen , S. H. Chua , D. Chung , S. T. Perrault , S. Zhao , and W. Kei . Eyes-free gesture passwords: a comparison of various eyes-free input methods . In Proceedings of the Second International Symposium of Chinese CHI , pages 89 { 92 . ACM, 2014 .

[8] G. Drosatos , P. S. Efraimidis , A. Arampatzis , G. Stamatelatos, and I. N. Athanasiadis . Pythia: A privacy-enhanced personalized contextual suggestion system for tourism . In Proc. of the 39th Annual IEEE Computer Software and Applications Conference (COMPSAC '15) , Taiwan, July, pages 1 {5 , 2015 .

[9] M. L. Dybdal , J. S. Agustin , and J. P. Hansen . Gaze input for mobile devices by dwell and gestures . In Proceedings of the Symposium on Eye Tracking Research and Applications , pages 225 { 228 . ACM, 2012 .

[10] D. Gavalas , C. Konstantopoulos , K. Mastakas , and G. Pantziou . Mobile recommender systems in tourism . Journal of Network and Computer Applications , 39 : 319 { 333 , 2014 .

[11] S. D. Gosling , P. J. Rentfrow , and W. B. Swann . A very brief measure of the big- ve personality domains . Journal of Research in personality, 37 ( 6 ): 504 { 528 , 2003 .

[12] K. Kularbphettong and B. Ngamkam . A recommendation system for heritage-tourism based on mobile application and ontology technique . International Journal of Information Processing and Management , 5 ( 3 ): 42 , 2014 .

[13] S. Naja an , W. Worndl, and B. Lamche . Investigation of user rating behavior depending on interaction methods on smartphones . In Joint Workshop on Interfaces and Human Decision Making in Recommender Systems , RecSys '14 conference, 2014 .

[14] I. Oakley and J. Park . Motion marking menus: An eyes-free approach to motion input for handheld devices . International Journal of Human-Computer Studies , 67 ( 6 ): 515 { 532 , 2009 .

[15] X. Page , B. P. Knijnenburg , and A. Kobsa . Fyi: communication style preferences underlie di erences in location-sharing adoption and usage . In Proceedings of the 2013 ACM international joint conference on Pervasive and ubiquitous computing , pages 153 { 162 . ACM, 2013 .

[16] P. Pawara , M. Kaenampornpan , and K. Kawattikul . Valuable tourism information via mobile application . In Advanced Materials Research , volume 1044 , pages 1428 { 1432 . Trans Tech Publ, 2014 .

[17] D. Rozado , T. Moreno , J. San Agustin, F. Rodriguez , and P. Varona . Controlling a smartphone using gaze gestures as the input mechanism . Human{Computer Interaction , 30 ( 1 ): 34 { 63 , 2015 .

[18] D. Sakamoto , T. Komatsu , and T. Igarashi . Voice augmented manipulation: using paralinguistic information to manipulate mobile devices . In Proceedings of the 15th international conference on Human-computer interaction with mobile devices and services , pages 69 { 78 . ACM, 2013 .

[19] J. Sodnik , C. Dicke , S. Tomazic , and M. Billinghurst . A user study of auditory versus visual interfaces for use while driving . International Journal of Human-Computer Studies , 66 ( 5 ): 318 { 332 , 2008 .

[20] G. Tumas and F. Ricci . Personalized mobile city transport advisory system . Information and Communication Technologies in Tourism 2009 , pages 173 { 183 , 2009 .

[21] M. Van Setten , S. Pokraev , and J. Koolwaaij . Context-aware recommendations in the mobile tourist application compass . In Adaptive hypermedia and adaptive web-based systems , pages 235 { 244 . Springer, 2004 .

[22] M. Zheng , O. Ormandjieva , and H. Fan . Designing context sensitive mobile user interface . In Proceedings of the International Conference on Software Engineering Research and Practice (SERP) , page 152. The Steering Committee of The World Congress in Computer Science , Computer Engineering and Applied Computing (WorldComp), 2015 .