Harmonizing divergent user preferences for cultural enrichment of small group visit

Abstract

People spend their leisure time on cultural and historical stories by visiting exhibitions and museums with family members and friends. This small gathering establishes a voluntary platform to engage in activities and share experiential knowledge. Since members of a small group tend to have diverse preferences, both individual and group preferences should be considered for a cohesive and harmonious guide. This paper proposes a mixed-initiative tour path planning system for multiple visitors that builds a group path supported by automatic path generation and users’ participation. In the proposed system, individual user profiles are merged into a group profile by favoring multiple users’ higher preferences with low deviation. Then the system constructs a tour path based on mixed-initiative interaction that consists of the system’s automatic decision and participation of members. When group members have similar preferences, a tour path is automatically decided by the system. If group members have different and divergent preferences, the system asks group members to intervene and confirm appropriate places to collaboratively construct a tour path. To show feasibility of our proposed system, we evaluate our approach in twofold consists of a synthesized group simulation and a user study. For the simulation, we collected user preferences of 12 real users to simulate groups of two, three, and four members and explored characteristics of similar and different preference scenarios. For the user study, we conducted a comparative study with 11 participants under 4 different exhibition scenarios using the implemented mixed-initiative path planning system. Evaluation results support that our system constructs a cohesive, satisfying and usable path for a small group tour.

Keywords

Collaborative tools human computer interaction path planning recommender systems cultural enrichment

1. Introduction

With recent advances in information and communications technology (ICT), a number of studies have attempted ICT-based solutions to improve tour guidance and user experience (UX) for cultural heritage (CH) visitors at museums and art galleries [4,14,21,28]. Museums equipped with such advanced technologies provide personalized multimedia information by integrating user-specific context with physical artifacts to help visitors better understand historical and cultural value of the exhibits. Museums are also a social place where groups of people share personal and experiential knowledge on the exhibits during their visits. Since people often visit museums with their family members and friends, there are increased needs and demands for supporting tour guidance targeting a small group of visitors [3,6,30].

Researchers previously have explored ICT-based approaches for improving tour guidance and UX for the context of museums. For example, earlier studies focused on providing personalized information based on location and time context. For these studies, social interaction between multiple users were mediated by personal handheld devices or their face-to-face communication [1,13]. Later studies then explicitly probed and incorporated cases of people visiting museums together in groups. A tour guidance system, Sotto Voce, allowed users to automatically share audio contents over its voice channel [3]. Similarly, another mobile tour system allowed users to join a community with their mobile devices to share audio content within that community [26]. Recent studies incorporated multi-user tabletop systems to support collaboration among group members. In these systems, group members were presented with information they can explore together [16] and manually selected visiting sites where they can use traditional social interaction (i.e, verbal communication) or collaboratively used the provided application [18]. Other group recommendation systems focused on automatic generation of multimedia contents. A recommendation system, MusicFX, automatically generated a list of music categories by aggregating individual preferences [20] while a TV program recommendation system generated a play list by merging individual profiles [33]. To sum up, abilities to share and collaborate among group members are crucial factors to improve UX in museums, yet there are only a few studies on improving tour paths for the group visitors. Although automatic generation of group recommendation is convenient, the resulting items unlikely satisfy group members due to their divergent preferences. In essence, members within a group often have different preferences and prior knowledge. Therefore, each member of the group is more likely to be interested in different exhibits or another tour path. Our motivations of tackling this imbalance are as follows.

In a group recommendation, both individual and group preference should be reflected in contrast to personal recommendation. However, since a group preference represents the integrated preferences of all members in the group, different and divergent individual preferences are often and inevitably neglected. Therefore, we need a group profile that can capture both individuals and the group aspects.

When a tour path is solely generated from a group profile, there often exists cases where the generated paths are not entirely satisfactory to all or partial members. In these circumstances, both automatic decision by the system and users participation should be encouraged. Then the users could finally reach and confirm a collaborative decision on the tour path.

Fig. 1.

Mixed-initiative path planning for a group.

To overcome aforementioned limitations, we propose a mixed-initiative path planning system for multiple visitors as shown in Fig. 1. The proposed system first merges users’ preferences to extract exhibits of common interest on a museum site and then builds a tour path for the group. To recommend a feasible path for the group, the system automatically selects places of common interest where the users have similar preferences. Otherwise (i.e., cases of divergent preferences), the system highlights those places of common interest and asks users to confirm remaining places they want to visit. Strategically, this approach allows each user to appeal for visiting places of personal interests as well as places of the group’s interests. We implemented our mixed-initiative path planning system on a tabletop setting targeting museums. To show feasibility of our system, we evaluated our approach in twofold consists of a synthesized group simulation and a comparative user study.

This paper is structured as follows. We first review related work on social interaction, group tour and group recommendation in Section 2. Then we describe our mixed-initiative path planning approach in Section 3. We further discuss results of our work in Section 4, Section 5 and Section 6 through implementation, evaluation and user study aspects, respectively. In Section 7, we summarize implications of our work followed by conclusion in Section 8.

2. Related work

2.1. Social interaction and group tour

A number of studies have been conducted for small group tours. There were earlier works on improving social experience of small group tours and visits. Aoki et al. proposed Sotto Voce, a mobile voice guidebook for a small group tour that allowed visitors to share each other’s voice contents [3]. If there was no active voice guide in one’s mobile guide, Sotto Voce allowed listening to a friend’s voice guide automatically. Through this automatic audio sharing, visitors became aware of other friend’s activity and thoroughly engaged in a cohesive group tour. Suh et al. also proposed a mobile guide system that supported voice sharing among multi-users [26]. Their mobile guide system allowed users to explicitly enable or disable sharing of voice contents. They found that friends in groups were more likely to finely control voice sharing than family members in groups.

Other works supported collaborative tour of small groups based on smart devices. Wakkary et al. proposed Kurio system consisted of mobile phones and a shared tabletop display for family group tours [29]. Each member was allowed to participate in a collaborative game-based tour and was able to see and share progress of the game through the tabletop display. By using mobile devices and shared display for collaboration, family members were engaged in their tour and game activities. Hope et al. observed group tour behaviors based on mobile guide systems, called Minparku Navi for family groups [13]. They found that visitors had different skills and expertise, and they were dependent on mobile devices as well as face-to-face communications. Klinkhammer et al. proposed a tabletop display to support collaboration of group members [16]. Users were allowed to not only manage their personal information, but also share them among members via the tabletop system. Marshall et al. proposed a tabletop display to support a path planning for group members [18]. With the presented tabletop setting, users were allowed to manually construct their path and share visiting information.

More recently, other studies have improved sharing capabilities for a cohesive group tour. Belinky et al. used both mobile devices and a shared display to share their visiting path in a museum [5]. The mobile devices allowed users to pre-plan places before their visit. The shared display allowed users to share their preferences and re-plan or modify their visits. Fosh et al. observed effects of group formation on group tours [10]. In their work, visitors were provided with different types of guide information and they were allowed to share their information with others. Through observation on friend and family groups, they found that group tours were frequently joined and separated by a group, and their asymmetric information encouraged the proactive collaboration among group members. Furthermore, Brown et al. proposed a mixed reality (MR) tour system that combined real, virtual and web-based museums for a small group [6]. This MR tour allowed visitors to be located in different spaces and enabled them to share and collaborate with other users in different spaces. Through the MR-based tourism, the visitors were able to understand and share different aspects of museums in each space.

Overall, previous work mainly focused on encouraging social interaction and sharing contents among group members where mobile devices and shared displays were typically used for collaboration. However, previous systems for small group tours did not consider guiding a group path for multi-users scenarios. To follow a group tour using such systems, visitors were required to frequently self-check and be aware of others’ activities to stay in the group tour. Still, visitors would face difficulties and lose track of the group tour when other members go out of their sight.

Table 1
A summary of representative related work

Reference Group size Application Aggregation function Decision making

Aoki et al. [3] 2 Group Tour None Automatic

Suh et al. [26] 2 Group Tour None Manual Selection

Wakkary et al. [29] 2 to 4 Group Tour None Face-to-Face

Hope et al. [13] 2 to 4 Group Tour None Face-to-Face

Klinkhammer et al. [16] 2 to 4 Information Sharing None Face-to-Face

Marshall et al. [18] 2 to 4 Group Tour None Face-to-Face

Belinky et al. [5] 2 to 4 Information Sharing None Face-to-Face

Fosh et al. [10] 2 Group Tour None Face-to-Face

Brown et al. [6] 2 to 3 Group Tour None Face-to-Face

McCarthy and Anagnost[20] 2 to 100 Music Play Average Automatic

Yu et al. [33] 2 to 4 TV Recommendation Average Automatic

Shin and Woo[24] 2 TV Recommendation Average Automatic

Garcia et al. [11] 2 to 4 Tour Recommendation Average & Most Popular Automatic

Shin et al. [23] 2 Media Control (TV, Music) Multiplicative & Average Automatic & Face-to-Face

Reference	Group size	Application	Aggregation function	Decision making
Aoki et al. [3]	2	Group Tour	None	Automatic
Suh et al. [26]	2	Group Tour	None	Manual Selection
Wakkary et al. [29]	2 to 4	Group Tour	None	Face-to-Face
Hope et al. [13]	2 to 4	Group Tour	None	Face-to-Face
Klinkhammer et al. [16]	2 to 4	Information Sharing	None	Face-to-Face
Marshall et al. [18]	2 to 4	Group Tour	None	Face-to-Face
Belinky et al. [5]	2 to 4	Information Sharing	None	Face-to-Face
Fosh et al. [10]	2	Group Tour	None	Face-to-Face
Brown et al. [6]	2 to 3	Group Tour	None	Face-to-Face
McCarthy and Anagnost[20]	2 to 100	Music Play	Average	Automatic
Yu et al. [33]	2 to 4	TV Recommendation	Average	Automatic
Shin and Woo[24]	2	TV Recommendation	Average	Automatic
Garcia et al. [11]	2 to 4	Tour Recommendation	Average & Most Popular	Automatic
Shin et al. [23]	2	Media Control (TV, Music)	Multiplicative & Average	Automatic & Face-to-Face

2.2. Group recommendation

Earlier work on group recommendation focused on recommendation of multimedia contents. McCarthy and Anagnost proposed MusicFX, a group music recommendation for users in a healthcare center [20]. In this work, they merged individual preferences based on the sum of square function and recommended music categories in the order of the preferences merged. Consequently, popular categories were recommended first followed by less popular categories. Yu et al. proposed a TV recommendation for a group of users [33]. Their approach merged a set of individual user’s profiles into a group profile and recommended TV programs based on the group profile. Shin and Woo proposed a TV recommendation using a similar group recommendation system [24]. Their approach merged user’s preferences with respect to personal, group and family to recommend related group items. Therefore, TV programs satisfying group characteristics were recommended at the end.

There were other work focusing on group recommendation for museum visits. Cosley et al. introduced ArtLinks, a social recommendation based on a collection of tags [7]. The visitors were allowed to add tags on artifacts. Then tag information related to the artifacts were collected to suggest a next similar item to visit. Garcia et al. proposed a group recommendation system for tourist activities [11]. Their system assigned higher priority when the majority of a group is involved and merged user profiles based on weight averages. Rossi et al. proposed a group path recommendation in museums [22]. To maximize group satisfaction, visiting places were selected based on multiplicative aggregation function and a group path was generated as a resource-constrained shortest path problem. As the results, the group path satisfying both the group and its members were automatically selected.

More recent work improved quality of group recommendation by using additional information of users and items. Zhu et al. proposed a context-aware group point-of-interests (POI) recommendation for mobile users [34]. This system utilized current location of users to rank candidate places to meet as a group. Toan et al. proposed a group recommendation system for diversifying group items [27]. This method combined a group aggregation function with a diversification algorithm where the diversification algorithm selected the best set of items having diverse preferences after aggregating individual preferences. Hong et al. exploited a concept of affinity from users’ artwork-watching histories for an artwork recommendation involving multiple users [12]. To calculate affinity between users, a similarity measure for artwork features and indirect artwork viewing experience are considered.

Overall, previous group recommendation studies mainly focused on automatic recommendation of group items. Indeed, introducing additional information can make group tours more satisfying for visitors. This approach is easy, convenient to use for visitors and applicable for cases when group members have similar preferences. However, in cases where many users have different and divergent preferences, members will not be satisfied by the quality of the recommended paths. In our work, we are more interested in cases of divergent preferences. For example, users with different preferences are forced to visit a museum as a group through the sole path that includes several places. In this case, there is a risk of employing the sole path, since the path may contain many uninteresting, boring and unsatisfying places for some or majority members of the group. To mitigate imbalance and improve overall satisfaction, we adopt multiplicative utility from group selection theory that takes care of both high preference and individual difference. We also incorporate a group confirmation consisting of recommendation and agreement step before finalizing the group path. A summary of some of the representative related work is presented in Table 1. The benefit of our work for the small group is that our approach is able to generate a group tour path for both cases of similar preferences and divergent preferences. Our approach is especially useful when members have some common interest and some conflicting interests. Our contribution is to use mixed-initiative approach to deal with users’ divergent preferences. In the later sections, we describe our system and its evaluation in detail.

Fig. 2.

The proposed path planning system with group negotiation.

3. Mixed-initiative path planning for small group tour

To deal with a group tour path generation, we incorporate the mixed-initiative approach that depends on both a computer system’s decision and user’s participation. In this approach, a computer system has the initiative to make a final decision when fixing a subtask or critical situation, otherwise, users are asked to participate in the decision-making process [2]. This approach is also useful in resolving conflicts of multiple users who share a smart space [23]. However, this approach presents several candidates that can be manually selected for a path planning where only a part of those candidates are eventually selected regarding preferences and time. Consequently, the mixed-initiative approach for the path planning should cover multiple items for multi-users. To do so, we exploit the system’s decision when the number of visiting items with different preference are rather small. When the number of visiting items with different preference are large, we prompt and invite visitors to select the remaining places together.

Next, we introduce our tabletop path planning system for a small group tour with the mixed-initiative interaction. The proposed path planning system generates a group profile by merging users’ individual profiles and builds a group path that contains places to visit. Afterwards, the system finalizes the group path based on the automatic decision-making and users participation. When the resulting path is acceptable and satisfies group members, the system should automatically confirm and declare the path as the final result. For other cases where users’ preferences diverge on the resulting path, the system should only recommend places to visit on a shared display, so that users can intervene and decide their final path.

For this purpose, our system consists of three modules: 1) Group Path Generation (GPG), 2) Group Path Management (GPM) and 3) Interaction and Visualization (InV), as depicted in Fig. 2. The InV module manages user profiles and map topology while being connected to a touchscreen device. This module also processes user inputs received from the touchscreen and visualizes maps and tour paths on the display. The GPM module manages users’ profile and tour information. This module feeds these data to GPG when multiple visitors start to build their path. The GPG module merges user profiles and then constructs an intermediate group path which includes places to visit for the group. This module also confirms the final group path aided by the system’s automatic decision and users participation. The right part of the Fig. 2 depicts the group negotiation process among the visitors with visual and interactive display. The visitors are allowed to see the intermediate group path indicating fixed, recommended and conflicting places. The fixed items are already selected while the recommended places are the candidates to be decided. The conflicting places are excluded and not allowed to be selected.

Based on these three modules, the proposed path planning system works as follows. The system first obtains individual profiles containing preferences on possible places to visit. Then the system normalizes and merges individual preferences for a group recommendation. Based on the merged profile, the system generates a tour path based on graph search algorithms over a museum map. The system then decides whether the generated path is acceptable according to satisfaction of group members over the included places to visit. When there is only a small number of places that have different and divergent preferences among group members, the system displays the generated path as the final result. Otherwise, the system recommends an intermediate path that contains places of different and divergent preferences. Then the system asks users to confirm places to visit through negotiation and selection among group members. Through these path generation and confirmation steps, the final tour path based on the system’s automatic decision and users’ negotiation is presented to users. In the following subsections, we present three components of the GPG and GPM modules in detail.

3.1. User profile merging

In any application involving multiple users, it is important to satisfy multi-users’ preferences. For this purpose, we generate a group profile by merging individual user profiles. We first define user profile, UP, as a collection of preferences ( $u p$ ) on visiting places (P) where preferences are represented as scalar ratings between 0 (not interested) and 10 (highly interested). The user profile is represented as ${U P}_{i} = ⟨ P_{1}, u p_{i 1} ⟩, ⟨ P_{2}, u p_{i 2} ⟩, \dots, ⟨ P_{N}, u p_{i N} ⟩$ where $P_{j}$ is the name of a visiting place or item, N is the number of places in a museum and $u p_{i j}$ is the preference of user i over $P_{j}$ . The group profile is then obtained from the user profiles. Since users have different preference ranges, we normalize users’ preferences based on minimum and maximum value of each user’s preference. Equation (1) illustrates the normalization of preferences in user profiles. Consequently, the normalized preference nup is obtained from original preferences of each user.

Afterwards, the normalized profiles are merged into a group profile. To reflect both individual preferences and misery of users, we adopt the multiplicative utilitarian from social selection theory [19]. Since multiplicative utilitarian is the multiplication of users’ preferences, highly rated items also score high. On the other hand, items with divergent preferences have relatively low scores. Equation (2) shows the equation for obtaining group profiles. As a result, group preference, gup is generated for the group. $\begin{array}{l} (1) & \begin{matrix} {nup}_{i, j} & = u p_{min} \\ + \frac{(u p_{max} - u p_{min}) \cdot (u p_{i, j} - u p_{i, min})}{u p_{i, max} - u p_{i, min}} \end{matrix} \\ (2) & {gup}_{j} = \prod_{i = 1}^{N} ({nup}_{i, j} + 1) \end{array}$

To illustrate Equation (1) and Equation (2), let us consider a case with two users in the group. User 1 has preference ratings between 4 and 6 while user 2 has ratings from 2 to 6. As mentioned earlier, the preference ratings are between 0 and 10. Now let us pick a place $P_{k}$ that both users will rate with a preference score of 5. By Equation (1), we have the normalized preference ${nup}_{1, k}$ for the place $0 + ((10 - 0) \cdot (5 - 4) / (6 - 4)) = 5$ , for user 1. For user 2, the normalized preference ${nup}_{2, k}$ is $0 + ((10 - 0) \cdot (5 - 2) / (6 - 2)) = 7.5$ . These results indicate that the preference score of 5 is a neutral score for user 1 while the same preference score of 5 leans toward relatively high preference for user 2 considering her ratings on other places. Finally, these preferences are merged by Equation (2) where the multiplication of the two items is calculated as $(7.5 + 1) \cdot (5 + 1) = 51$ .

3.2. Group path generation

After the group profile is obtained, next step is to generate a tour path for a museum site. Since there are many places to visit in the museum, the path generation should select a set of places with associated visiting time and generate the shortest path for the tour. Additionally, selection criteria should consider both individual and group preferences to satisfy group members. Therefore, we consider this path selection as a selective traveling salesman problem (STSP) that selects probable places from start location to destination within a given time [17]. We cannot travel all possible alternatives from a traveling map, because STSP have so many solutions. Therefore, we approximate the best solution. In our work, we selectively increase visiting places and then build an alternative path altogether. Unlike travel itineraries [31] or theme park tourist services [32], we intentionally neglect the time to move between places, because moving time is short compared to the actual staying time (i.e., time spent to appreciate the beauty of a work of art) in indoor museums.

Algorithm 1

Algorithm for group path generation

Algorithm 1 shows how the visiting places are selected and the tour path is generated. Visiting list $T$ represents consecutive visiting places, $P_{S}$ , …, $P_{D}$ . $L (T)$ is the time required for visiting places in the list while c is the given time for the group. $Path (P_{S}, P_{D})$ is a function generating the shortest path from $P_{S}$ to $P_{D}$ based on A* search. In our work, we are interested in selecting visiting place and its connection because of limited tour activity time. A* is the efficient path generation algorithm used in various area such as path finding and navigation. It requires smaller memory and produce fast results. As a heuristic function, an estimated Euclidean distance in terms of x and y coordinates to the final destination (i.e., exit) is used. We used the A* algorithm to incrementally build a path by adding or selecting a visiting place by users or the system. Thus, the algorithm produces intermediate path fast whenever users add or remove a visiting place during path mediate process. The presented algorithm works as follows. Initially, the algorithm selects a start place $P_{S}$ and a last place $P_{D}$ . The algorithm finds the best place $P_{j}$ having the highest group preference and then adds that place to the visiting list. This selection step repeats while the total visiting time is less than the given visiting time. Finally, the algorithm produces the visiting list that maximizes the group preference.

3.3. Path confirmation in group path management

Although places to visit are included in the intermediate group path, the path is not finalized yet. Since users have different preferences on the selected places, users may not agree on the list. Here we introduce a path confirmation step in the GPM layer to negotiate the intermediate path. To make the final path agreeable and cohesive, we utilize mixed-initiative interaction approach. Specifically, path planning by the system and users participation are combined according to users’ preferences with respect to a group preference. When users’ preferences are similar, the path planning system can safely select the intermediate path as the final. Otherwise, the path planning system only recommends candidate places to visit and asks users to confirm their visiting places. For this purpose, we measure preference divergence between users and count the number of places of conflicts ( $P_{conf}$ ) whose preference divergence exceeds a certain threshold d. When the number of such places with divergent preferences, is greater than a certain threshold $T_{grp}$ , this group is classified as a different-preference group (Diff-Pref), otherwise the group consists of users with similar-preferences (Sim-Pref). Equation (3) shows how a group is classified into Diff-Pref or Sim-Pref group according to preference divergence. Here $P_{sel}$ is the place selected in the intermediate path and $P_{conf}$ is the place that should be decided later. $Diverge (P_{i k})$ is the preference divergence of $P_{i k}$ (i.e., the preference of user i over $P_{k}$ ). Therefore when the number of $| P_{conf} |$ is greater than or equal to $T_{grp}$ , the group is considered as a different-preference group. $\begin{matrix} (3) & Pref = \{\begin{matrix} Sim-Pref \\ if | P_{conf} | < T_{grp}, \\ where P_{conf} = {P_{i k} | P_{i k} \in P_{sel} \\ and Diverge (P_{i k}) < d \\ for every user i over P_{k}} \\ Diff-Pref \\ otherwise \end{matrix} \end{matrix}$

With this intermediate result, path confirmation step starts in the GPM. During the path confirmation, $P_{conf}$ are highlighted and selected by users while other places are being fixed. The next path is updated whenever the users confirm $P_{conf}$ . The final path is selected after this visualization followed by users’ negotiation and selection.

Fig. 3.

A working demonstration of the implemented tabletop path planning system.

4. Implementation

We implemented the proposed path planning system for simulating and validating our path planning approach. To provide easy access to multi-users, the proposed system was implemented as a tabletop system where the top surface contained a large touchscreen. Underneath the tabletop, a desktop computer was installed. We implemented the path planning system using Microsoft Visual Studio. Open Scene Graph was used for the visualization and interaction module. Figure 3(a) shows examples of a topology and a generated tour path. Figure 3(b) shows the implemented path planning system on a tabletop. We presented a graphical user interface (GUI) to visualize a map and visiting places and to get user inputs for path confirmation. As depicted in Fig. 3(b), the path planning system included four buttons to select maps and two buttons to add or delete places to visit. It also visualized places with similar preference or already determined places in green. For other places, they were highlighted with red whose preferences of group members were different and divergent. The lines between visiting places indicated the shortest path from the start to destination traversing the visiting places. With this tabletop system and supported visualization, users were automatically provided with a recommended path to visit a museum when they had similar preferences on visiting places. They had an option to accept and use the suggested path or they could further review and change visiting places for their tour.

5. Evaluation of group path recommendation

5.1. Evaluation setup

We first analyzed the quality and characteristics of group path recommendation and divergent preferences. For this purpose, we prepared a virtual museum as shown in Fig. 3(b) that mimics real museums and contains four scenarios with different themes (e.g. history, nature, culture). Each scenario included 10 visiting places with a set of related exhibits. We then recruited 12 participants and they were asked to rate their preferences on these places. We then simulated three sizes of small groups for group recommendation (i.e., groups of two, three and four members, respectively). We randomly selected two, three and four users from the 12 participants. We obtained 100 synthesized groups for each two, three, four user groups. Using these groups, we normalized and merged user preferences. Then we obtained recommended places for the groups and analyzed divergent preferences of the recommended places for users in that group.

We also compared the quality of aggregation function used in group profile merging. In previous work, group preference is frequently generated by calculating an average of user preferences [11,20,24,33]. Therefore we compared recommendation quality between average utility and multiplicative utility functions on the group preference generation stage. For this purpose, as recommendation quality metrics, we calculated and compared average preferences and preference deviation.

5.2. Result

We found that occurrences of the divergent user preferences increased with the number of recommended places. As seen in Fig. 4, the preference divergence of two users group (GP2), three users group (GP3), and four users group (GP4) were 1.4, 1.8, and 1.9, respectively when only one place was recommended. The computed divergence value reached its maximum when the number of recommendation places was 7. The divergence value then decreased beyond that point, because user preferences were very low when 8 to 10 places were recommended. To summarize, preference divergence was low when there were only small number of places to recommend. The preference divergence was increased as the number of recommend places increased, until preferences of each place became low.

Fig. 4.

Preference deviation simulation of group sizes of two, three, and four members.

We also found that the number of places with different preferences, as well as the number of places with common preferences were increased. For this purpose, we focused on the common and different places of group recommendation in two user groups. The common places were places automatically recommended by the group recommendation which include users’ preferred places. On the other hand, different places were places where the preference divergence between two users was higher than a certain value (i.e., 4 for our experiment).

Fig. 5.

Common and different places according to the number of recommendation places.

As seen in Fig. 5, the average number of the common places increased from 0.5 to 10 while the average number of different places increased from 0.1 to 4. To summarize, if the number of recommended places was small, the number of places with divergent preference was small and the preference divergence was also low. Due to the fact that places of different preferences were increased as more items were recommended, the preference divergence also increased inevitably. Arguably, to plan for a group tour with a small number of participants and to deal with divergent preferences, both individual and group preferences should be considered in the planning system.

Fig. 6.

Comparison of multiplicative utility and average utility.

We found that the multiplicative utility function used in our path planning system is better than the average utility function in terms of preference deviation. As seen in Fig. 6(a), Top-5 items generated from the multiplicative utility function has low preference deviation compared to the items generated from the average utility function. The preference deviation of multiplicative utility and average utility was 1.7 and 1.8 respectively when only 1 item is recommended. This gap was widened up to top-5 item recommendation. The average of the difference of top items of the multiplicative utility and average utility was 3.5 and 3.9, respectively. In spite of such difference, the average preference of the two utility functions was similar as shown Fig. 6(b). Therefore it is clear that the multiplicative utility function is useful for yielding the best items for group users by selecting items of high preference and low deviation. In next section, we report on satisfaction of users on the group path planning in terms of decision-making.

6. User study

6.1. Procedure

We conducted a pilot study to evaluate quality of the path confirmation step with participants.1

¹
A partial result has been appeared in a Korean-language paper [25] intended for Korean readers.

In previous studies, paths were mainly confirmed by their systems based on the system-produced recommendation results [11,20,24,33]. In other words, they assumed that the group recommendation is already final. In this work [23], they only reflected the user’s participation during decision making. To evaluate this step, we simulated the path confirmation step combined with the automatic decision making and the user’s participation. We simulated path planning scenarios and asked users to participate in path planning activities using the virtual museum depicted in Fig. 3(b). For this purpose, we prepared 4 map scenarios with different themes to represent diversity of museums and user preferences. We then recruited 11 participants aged from 20 to 35. They were graduate students who had previously visited museums together with their family members or friends. We then asked each participant to rate his/her preferences on the places on example maps. The maps included history, drawing, person and science sections. There were 10 places in each of the map scenarios and preference rating was an 11-level likert scale (0 to 10). Afterwards, the rated preferences were saved in the system for further path recommendation. Before our main study, each participant had a practice session to understand how to use and control the path planning system. This part of our study was based on Wizard of Oz [8], so we introduced an actor who was acting as a member of the participant’s group. We asked the participant to familiarize himself/herself with the actor and practice a decision-making together.

In the main study, each participant experienced and evaluated 4 scenarios with different preferences and confirmation steps. To support 4 scenarios, we combined two preference scenarios of similar and different categories with two path decision strategies of automatic path decision and users confirmation. We focused on selecting 5 places from 10 places in the decision making process of each scenario. For this purpose, we set $T_{grp}$ as 3 while the divergence threshold d of 4 was used to divide the types of preferences according to Fig. 4(b). Consequently, a number of included places with different preferences varied, for each similar-preference and different-preference scenario. For the similar-preference scenario, we simulated preferences of the actor to have small divergent preference value less than 3. The path planning system decided top 5 places. In automatic decision, 5 places with higher preferences were selected. In path confirmation, participants were asked to select one of two recommended places together with the actor, in addition to four fixed places. For the different-preference scenario, we simulated preferences to have larger divergence values distributed from 1 to 7. In this situation, automatic decision selected 5 places while path confirmation asked the participant to select 3 of 6 recommended places in addition to two fixed places. Based on this setting, participants experienced both automatic path selection and path confirmation in similar and different preference situations. Soon after finishing each scenario, we asked them to rate the quality of path recommendation in terms of overall satisfaction and five factors such as understanding of the path planning procedure (U), controllability over the results (C), appropriateness of the results (B), preference reflection (P) and agreement on the solution (A). The level of each quality was 5-level likert scale. Finally, we asked the participants to describe their thoughts on the path planning and path confirmation process in a short interview session.

6.2. Results

We found that the participant preferred the automatic path planning when their preferences were similar. However they preferred to be involved in path confirmation when their preferences were different. As seen in Fig. 7, overall satisfaction of the automatic decision (AD) and user confirmation (UC) were almost the same when their preferences were similar (i.e., cases of low divergence values). However, satisfaction on the UC was superior to that of the AD when their preferences were different (i.e., cases of large divergence values). The satisfaction of the UC slightly decreased in large divergent cases, while satisfaction on the AD largely and statistically decreased, $t (10) = 2.49$ , $p = 0.032$ .

Fig. 7.

User satisfaction on path planning strategies.

Fig. 8.

Satisfaction on five factors according to path planning strategies (a) similar preference, (b) different preference.

When we closely look into five factors of satisfaction as shown in Fig. 8, we also found that the participants were highly satisfied with both AD and UC when they had similar preferences. However, the perceived qualities of AD and UC were different when their preference were different. Although the resulting path reflected their preferences partially and were appropriate for the group, they were not fully satisfied with other factors when their preferences were different. Participants reported that the AD did not perform as well as the UC. They thought they did not have control over the selection (C, $t (10) = 4.38$ , $p = 0.001$ ), did not agreed with the selection (A, $t (10) = 2.80$ , $p = 0.02$ ) and did not understand well on how the path was selected (U, $t (10) = 2.86$ , $p = 0.02$ ). In spite of this unsatisfaction, they thought that such solution was inevitable for the group.

Lastly, we collected a wide range of opinions about the path planning strategies for small groups from individual interviews. Overall, participants were satisfied with the AD when their resulting path includes the majority of their preferred places. In this case, they tended to tolerate small part of recommended items with divergent preferences. They even described that the UC step did not changed the visiting path much in the similar preference scenario. A couple of the participants mentioned that simply changing one or two items produced better visiting plan for the whole group. However, participants had diverse opinions about path planning strategies for different preferences cases. Some participants thought that automatically selected places were what they had expected. They were actively involved in the UC step and wanted to change items in the path. During the UC step, most participants understood which visiting places were selected and which places can be changed. Many participants described that they planned to get best suitable items for the group tour during the UC while just one participant expressed that she explicitly planned to get her best places.

7. Implications

Users participation reduces the diverged preference gap and makes the solution more acceptable. In group recommendation where diverse preferences are involved, it is imperative to deal with different and divergent preferences of users. Therefore, a group recommendation algorithm should acknowledge and produce solutions by mitigating the preference gap problem. Through the additional user participation step to the path planning, users become aware of the group’s common and different preferences to understand each other. Furthermore, active discussion and negotiation allow them to share their desire and enable them to make an agreement on a better solution. For the group visit, the imperative preference divergence of group members is mitigated during this process.

Path planning should be automatic as much as possible. Although users participation contributes to make an agreed solution for the group, a path planning system should automatically produce the suitable path as much as possible. The system should automatically generate a proper path for the group when users have similar preferences. For other cases, the system could automatically select in parts or ask users to select items. In our work, the proposed system automatically selects items with small preference divergence and recommends an intermediate path while the system gets confirmation on items with highly divergent preferences. This structure relies on the automated part of the path planning and only exploits manual item selection by user participation on a need-to-basis.

Both individual and group preferences are important for satisfying all members of a group. As path planning is a problem of visiting places, it is important to reflect both individual and group preferences. Although the group preference is made from merging individual preferences, it seldom reflects divergent preferences of group members. When the majority of group members like places included in a tour path, still a small number of members who dislike those places is not satisfied with the suggested path. To solve this situation, path planning systems should extract highly preferred items for the group and deal with largely divergent items accordingly. In our path planning system, the multiplicative utilitarian function gives higher priority to items with similar preferences and higher preferences. Further, users are asked to make an agreement to add or remove places with divergent preferences in the path confirmation step.

There are several limitations in our proposed approach. Firstly, gathering user preference should be automated as much as possible. Although users have no or partial knowledge about exhibition and artifacts, their preferences on the artifacts or exhibit items are important for building their path. In our work, users are asked to set their preference with partial knowledge about the visiting places. However it is cumbersome to rate every artifacts before visiting. Through our observation, we found that users were hesitant to set their preferences with partial knowledge and also spent little time for preference setting. To solve this limitation, the path planning system should include automatic preference generation methods by exploiting data mining or collaborating filtering methods. In data mining approaches, user preferences are automatically predicted by their personal information [2] while collaborative filtering methods find similar users to estimate a new user’s preference [9]. Otherwise, mobile devices could also offer users’ preferences via network before group path planning starts.

Secondly, the museum map used in the path planning should be more realistic. In our work, we simply modeled the museum containing relatively small areas for validating path recommendation and path confirmation. Generally, museums and exhibitions include more complex spaces with multiple floors and sections. Museums also provide more information about exhibits and artifacts for visitors. For dealing with these issues, future path planning systems should support larger areas containing a number of visiting areas and POIs with enriched semantic and ontology-based models [15]. Moreover, the planning systems should integrate more interactive and intuitive methods to control map and visiting paths. With these improvements, an empirical study should be conducted to gather users’ feedback and experience involving diverse small groups.

Lastly, our experimental results only apply to and can be interpreted in the context of a small group tour. We have shown that when the path planning algorithm constructs paths with a higher number of places, there is increased occurrences of divergent preferences in the group. Therefore, a small group tour planning system could rely on the prompt menu and let the users manually decide the tour path. However, we cannot say that our approach applies or is effective to a planning system that scales beyond our tested conditions (i.e., a large number of places with highly divergent preferences in a more sizable group). For such scaled-up and complicated scenarios, we would need more advanced automatic decision-making approach, since a mixed-initiative approach for a larger audience is much costly and becomes impractical. Nonetheless, our proposed system introduced a practical solution for small group tour problems that expose interesting and practical use cases for integrating automatic, manual and mixed-initiative approaches.

8. Conclusion

In this paper, we proposed a mixed-initiative path planning system for a small group such as family or friends (i.e., up to 4 members) visit. The proposed path planning system extracted a group profile by merging individual profiles and then generated an alternative path for a group. With the intermediate path, the proposed system automatically finalized the path when the preference of users were similar, but asked users to decide their path when their preferences were different and divergent. We implemented and evaluated our proposed path planning system with a synthesized group simulation and a comparative user study. Through the synthesized group simulation, we found that the preference divergence increased as the number of recommended places in the group path increased. Through the user study, we found that users preferred the automatic path planning when they had similar preferences. In contrast, they preferred the path planning with the user confirmation step when they had different and divergent preferences. Consequently, it is strategically meaningful and useful for path planning systems to support both automatic decision and users participation for satisfying members in a group with divergent preferences.

The presented work is an initial step for developing mixed-initiative group path planning systems and there are several remaining open issues. For future work, we would like to extend the path planning system for more realistic environment enriched by augmented reality (AR) and virtual reality (VR). We are also interested in making the path planning system more interactive and efficient with cross-device user interfaces. To make path planning experience seamless and more satisfiable, users preferences available in various smart devices such as wearable devices, wearable glasses, and head mounted displays will be obtained, integrated and shared.

Footnotes

Acknowledgements

This work was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (NRF-2018R1D1A1B07043983). The authors thank the anonymous reviewers for their feedback and suggestions and acknowledge a Korean-language paper [] intended for Korean readers as an initial basis of this work.

References

G.D.

Abowd,

C.G.

Atkeson,

Hong,

Long,

Kooper and

Pinkerton, Cyberguide: A mobile context-aware tour guide, Wirel Netw 3(5) (1997), 421–433. doi:10.1023/A:1019194325861.

Allen,

C.I.

Guinn and

Horvitz, Mixed-initiative interaction, IEEE Intelligent Systems and their Applications 14(5) (1999), 14–23. doi:10.1109/5254.796083.

P.M.

Aoki,

R.E.

Grinter,

Hurst,

M.H.

Szymanski,

J.D.

Thornton and

Woodruff, Sotto voce: Exploring the interplay of conversation and mobile audio spaces, in: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, 2002, pp. 431–438.

Bartolini,

Moscato,

R.G.

Pensa,

Penta,

Picariello,

Sansone and

M.L.

Sapino, Recommending multimedia visiting paths in cultural heritage applications, Multimedia Tools and Applications 75 (2016), 3813–3842. doi:10.1007/s11042-014-2062-7.

Belinky,

Lanir and

Kuflik, Using handheld devices and situated displays for collaborative planning of a museum visit, in: Proceedings of the 2012 International Symposium on Pervasive Displays, 2012, p. 19.

Brown,

MacColl,

Chalmers,

Galani,

Randell and

Steed, Lessons from the lighthouse: Collaboration in a shared mixed reality system, in: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, 2003, pp. 577–584.

Cosley,

Lewenstein,

Herman,

Holloway,

Baxter,

Nomura,

Boehner and

Gay, Artlinks: Fostering social awareness and reflection in museums, in: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, 2008, pp. 403–412.

Dow,

MacIntyre,

Lee,

Oezbek,

J.D.

Bolter and

Gandy, Wizard of oz support throughout an iterative design process, IEEE Pervasive Comput 4(4) (2005), 18–26. doi:10.1109/MPRV.2005.93.

M.D.

Ekstrand,

J.T.

Riedl and

J.A.

Konstan, Collaborative filtering recommender systems, Foundations and Trends in Human–Computer Interaction 4(2) (2011), 81–173. doi:10.1561/1100000009.

10.

Fosh,

Benford and

Koleva, Supporting group coherence in a museum visit, in: Proceedings of the 19th ACM Conference on Computer-Supported Cooperative Work & Social Computing, 2016, pp. 1–12.

11.

Garcia,

Sebastia and

Onaindia, On the design of individual and group recommender systems for tourism, Expert Syst Appl 38(6) (2011), 7683–7692. doi:10.1016/j.eswa.2010.12.143.

12.

Hong,

Jung,

Piccialli and

Chianese, Social recommendation service for cultural heritage, Pers Ubiquit Comput 21(2) (2017), 191–201. doi:10.1007/s00779-016-0985-x.

13.

Hope,

Nakamura,

Takahashi,

Nobayashi,

Fukuoka,

Hamasaki and

Nishimura, Familial collaborations in a museum, in: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, 2009, pp. 1963–1972.

14.

Kesorn,

Juraphanthong and

Salaiwarakul, Personalized attraction recommendation system for tourists through check-in data, IEEE Access 5 (2017), 26703–26721. doi:10.1109/ACCESS.2017.2778293.

15.

Kim,

Matuszka,

J.I.

Kim,

Kim and

Woo, Ontology-based mobile augmented reality in cultural heritage sites: Information modeling and user study, Multimedia Tools and Applications 76(24) (2017), 26001–26029. doi:10.1007/s11042-017-4868-6.

16.

Klinkhammer,

Nitsche,

Specht and

Reiterer, Adaptive personal territories for co-located tabletop interaction in a museum setting, in: Proceedings of the ACM International Conference on Interactive Tabletops and Surfaces, 2011, pp. 107–110. doi:10.1145/2076354.2076375.

17.

Laporte and

Martello, The selective travelling salesman problem, Discrete applied mathematics 26 (1990), 193–207. doi:10.1016/0166-218X(90)90100-Q.

18.

Marshall,

Morris,

Rogers,

Kreitmayer and

Davies, Rethinking ‘multi-user’: An in-the-wild study of how groups approach a walk-up-and-use tabletop interface, in: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, 2011, pp. 3033–3042.

19.

Masthoff, Group modeling: Selecting a sequence of television items to suit a group of viewers, in: Personalized Digital Television, 2004, pp. 93–141. doi:10.1007/1-4020-2164-X_5.

20.

J.F.

McCarthy and

T.D.

Anagnost, Musicfx: An arbiter of group preferences for computer supported collaborative workouts, in: Proceedings of the 1998 ACM Conference on Computer Supported Cooperative Work, 1998, pp. 363–372.

21.

Pavlidis, Recommender systems, cultural heritage applications, and the way forward, Journal of Cultural Heritage 35 (2019), 183–196. doi:10.1016/j.culher.2018.06.003.

22.

Rossi,

Barile,

Galdi and

Russo, Recommendation in museums: Paths, sequences, and group satisfaction maximization, Multimed Tools Appl 76(24) (2017), 26031–26055. doi:10.1007/s11042-017-4869-5.

23.

Shin,

A.K.

Dey and

Woo, Mixed-initiative conflict resolution for context-aware applications, in: Proceedings of the 10th International Conference on Ubiquitous Computing, 2008, pp. 262–271.

24.

Shin and

Woo, Socially aware tv program recommender for multiple viewers, IEEE Trans Consum Electron 55(2) (2009), 927–932. doi:10.1109/TCE.2009.5174476.

25.

Shin and

Woo, Hybrid group path planning system for multiple visitors, Journal of the HCI Society of Korea 5(2) (2010), 25–31, (in Korean). doi:10.17210/jhsk.2010.11.5.2.25.

26.

Suh,

Shin,

Woo,

Dow and

MacIntyre, Enhancing and evaluating users’ social experience with a mobile phone guide applied to cultural heritage, Pers Ubiquit Comput 15(6) (2011), 649–665. doi:10.1007/s00779-010-0344-2.

27.

N.T.

Toan,

P.T.

Cong,

N.T.

Tam,

N.Q.V.

Hung and

Stantic, Diversifying group recommendation, IEEE Access 6 (2018), 17776–17786. doi:10.1109/ACCESS.2018.2815740.

28.

Varini,

Serra and

Gucchiara, Personalized egocentric video summarization of cultural tour on user preferences input, IEEE Transactions on Multimedia 19 (2017), 2832–2845. doi:10.1109/TMM.2017.2705915.

29.

Wakkary,

Hatala,

Muise,

Tanenbaum,

Corness,

Mohabbati and

Budd, Kurio: A museum guide for families, in: Proceedings of the 3rd International Conference on Tangible and Embedded Interaction, 2009, pp. 215–222.

30.

Wakkary,

Muise,

Tanenbaum,

Hatala and

Kornfeld, Situating approaches to interactive museum guides, Museum Management and Curatorship 23(4) (2008), 367–383. doi:10.1080/09647770802517423.

31.

Yoon,

Zheng,

Xie and

Woo, Social itinerary recommendation from user-generated digital trails, Pers Ubiquit Comput 16(5) (2012), 469–484. doi:10.1007/s00779-011-0419-8.

32.

F.C.

Yu,

P.C.

Lee,

P.H.

Ku and

S.S.

Wang, A theme park tourist service system with a personalized recommendation strategy, Appl Sci 8 (2018), 1745. doi:10.3390/app8101745.

33.

Yu,

Zhou,

Hao and

Gu, Tv program recommendation for multiple viewers based on user profile merging, User Model User-adapt Interact 16(1) (2006), 63–82. doi:10.1007/s11257-006-9005-6.

34.

Zhu,

Wang,

Cheng,

Sun,

Yang and

R.N.

Chang, Context-aware group recommendation for point-of-interests, IEEE Access 6 (2018), 12129–12144. doi:10.1109/ACCESS.2018.2805701.