The ability of users to organize events using mobile devices is a defining characteristic of today’s social network systems. With the advancement of mobile technology, more and more people are digitally connected, which makes the analysis of group event planning process and decision making group behavior and suggestions for event organizers critically important. There is a rich history of UbiComp research concentrating on individual user behavior analysis, which treats individual user’s data as a singleton. However, social interactions among group members are often ignored, and relatively scant research to date has explored the subject of group event scheduling mobile group dynamics. Some early work (Beckmann and Gross, 2011; Park et al., 2008) has been confined to in-lab surveys and has not studied the real-world event scheduling process bydynamics of mobile groups of users, nor the factors that would its impact on group event decision-making. More recently, a study of university groups using mobile phones was presented (Jayarajah et al., 2015). What has been missing to date is a detailed understanding of the process dynamics of how groups make a decision to visit a particular place at a particular time using their mobile devices. What factors influence a group’s final decision? This paper provides detailed novel insights in the event scheduling decision-making process of social mobile groups. We believe that this is an exciting area ripe for exploration by the ubiquitous computing research community. With ever-increasing popularity of smartphones, we expect that mobile computing will be used extensively to assist groups of people in event planning making decisions in terms ofabout when and where to rendezvous. Consider a group of friends out on a weekend evening trying to decide what movie to see or where to eat, or consider a group of professional colleagues trying to decide where to go for lunch. Given the frequency with which people schedule colocated events, we believe mobile applications for group event scheduling can provide significant help.
However, despite this considerable potential, today’s technology offers limited help when it comes to coordinating group events in online and offline scenarios. Currently there are few group event organization applications on the market. The most commonly used services are Meetup (meetup, 2017), Facebook Events (facebook, 2017) and Evite (Evite, 2017). In these services, hosts organize offline events and post them on the website. Users or group members who are interested in these events RSVP and later attend the events in real life. However, in all of these existing services, meeting time and location are settled by the host at the creation of the event. Potential users are not able to sufficiently express their opinions on when and where to meet. This will more or less have a negative impact on event attendance. Doodle (Doodle, 2017) is an online event scheduling service which supports groups in finding a mutually agreeable meeting time. Participants are able to vote for their time preferences. But all the meeting time options are pre-selected by the host added by host in the beginning. Group members have no permission to suggest new options. In addition, group members cannot suggest or vote for meeting locations.
To address the limitations of existing services, we developed OutWithFriendz, a mobile application that enables groups of people to decide together through a voting process the date/time the group would like to meet as well as the location where they would like to meet. OutWithFriendz is implemented as a client-server architecture that is comprised of both iOS and Android based clients that communicate with a server implemented as a Java Web application.
The main elements of our OutWithFriendz mobile application are shown in Figure 1. To start using it, a user may create a new invitation acting as a host. During this process, she can specify the details of this invitation including a title, a list of suggested dates, a list of suggested locations and invited participants. After this host submits a new invitation to the server, all invited participants receive it and can view the detailed invitation information on their own clients. They can then suggest more dates, locations or vote for their preferred options and comment on the invitation. After the voting process has ended, the host then decides on the final meeting location and time, whuch are then sent to all participants. In this example (Figure 1), the host suggested four locations and three date/time options. After the suggesting and voting process, she selected Cafe Mexicali and Friday, 03-17-2017, 18:00 as final decisions, which received the most votes. Please note that, in our design, the host can make decisions based on the voting results, but she does not have to always obey them. We do find few hosts in our field study whose final decisions was different from the options that received most votes. This scenario will be discussed later.
Introducing our newly designed OutWithFriendz mobile application which embeds group decision making into the voting process raises new questions: How do the mobile app users collaborate to organize their group events? What are the major factors that will impact group decisions? How is the voting behavior processed? And how to improve the event attendance rate?
Our mobile application also collects user mobility-related data. The app posts GPS user location traces to the server. Users may opt out from providing their location traces, although most users did not disable location tracking for the entire duration of their participation in the study. This user mobility data provides great opportunity to derive such input factors as spread, movement, mobility, and to investigate their impact on group event scheduling decisions.
The contributions of this paper can be summarized as follows:
The paper describes the design and implementation of a new mobile group application for group event scheduling and its supporting system: OutWithFriendz provides smooth functions for group hosts to easily create an invitation and invite other users to join. All group members can suggest their preferred meeting locations and date/time, ideal locations, meeting dates and vote for them. The host finalizes meeting location and time based on the voting results.
The OutWithFriendz system represents the first field-based study of the group event scheduling and decision-making process event organization behavior in the context of a deployed mobile application with widespread geographic usage. This has allowed us provides the possibility to collect precise user trace traces data.
Using the data collected from a field study of this novel system, we discovered a series of factors, such as mobility, host preference, user preference, and social voting influence that are significant in group event planning and decision making processes. A correlation analysis of these factors is also performed. Our study offers new insights for group hosts and members to improve their real-life event organization.
The rest of this paper is organized as follows: After discussing the related works, we introduce our system design and data collection. Next we present our data analysis and correlation analysis in more detail. Finally, we summarize the important results, highlight the key findings, discuss their potential usage the limitations, and conclude this work.
2. Related Work
In this section, we discuss works that are most relevant to ours. We divide them into three categories.
Group behavior analysis has been an area of active research in recent years. Li et al. designed a temporal-spatial method for group detection, locating and tracking (li2016temporal). Sen et al. designed a group monitoring system for urban spaces (Sen et al., 2014), and Jayarajah et al. studied how users’ mobility patterns change when they are within a group (Jayarajah et al., 2015). The work by Jamil et al. proposed a hybrid sensing system for monitoring crowd dynamics in large-scale events such as large religious gatherings (jamil2015hybrid). Brown et al. investigated the differences between individual and group behavior with respect to physical location (Brown et al., 2014). However, previous works have limited understanding of how groups make decisions and the influence of the group host and participants. Our system allows users to easily express their preferences through suggesting and voting for meeting locations and times. Analyzing the decision making process data collected within our system can provide insights that enable groups to better organize events.
Group event organization has been studied in a number of recent works. Yu et al. proposed a Credit Distribution-User Influence Preference algorithm to recommend potential participants to the group host (Yu et al., 2015). Zhang et al. and Pramanik et al. built models to predict event success in Meetup (Zhang and Lv, 2017; Pramanik et al., 2016). The work by Du et al. collected a series of contextual factors to predict individuals’ activity attendance (Du et al., 2014). However, all of these works collected their datasets from large public social websites such as Meetup and Douban. In these services, the group host makes the event decision on their own. The meeting location and time are settled when the invitation is created. An interactive group event scheduling decision making process is missing in these services.
Social influence occurs when an individual’s emotions, opinions, or behaviors are affected by others. This phenomena has been observed in many domains. For instance, Goyal et al. proposed an influence maximization method based on a historical user action log (Goyal et al., 2011, 2010). Li et al. further incorporated friend and foe relations in social networks for influence maximization (Li et al., 2013). Using a Meetup dataset, Zhang et al. demonstrate the significant impact of group leaders in making event decisions (Zhang and Lv, 2017). In the Doodle voting application, Zou et al. find that in open polls, the voting decisions by later respondents are highly influenced by early and nearby respondents (Zou et al., 2015). In this paper, we analyze the difference between the impact of the group host and participants on group event planning decision making, as well as the behavior of early and late-coming voters, and how early voters affect late voters, which to the best of our knowledge has not been previously studied.
User mobility analysis has been an area of active research in recent years. In previous work, human mobility patterns are analyzed to identify taxi trips (ganti2013inferring), to predict user’s future movement (baumann2015population), and to monitor whether individuals are affected by depressive mood disorders (canzian2015trajectories). No existing work investigates the influence of human mobility on group event attendance. Our study based on the dataset collected by our mobile application is valuable for understanding this effect. Our extensive analysis will show that user mobility prior to the group activity has significant impact on the user’s availability on the group event planing process.
Our work is also generally related to group behavior analysis. Sen et al. designed a group monitoring system for urban spaces (Sen et al., 2014), and Jayarajah et al. studied how users’ mobility patterns change when they are within a group (Jayarajah et al., 2015). Lampinen et al. investigated how users deal with group co-presence to prevent conflictive situations (Lampinen et al., 2009). There are also works studying group colocated interactions using mobile devices (Fischer et al., 2013; Lucero et al., 2015). The work by Brown et al. investigated the differences between individual and group behavior with respect to physical locations (Brown et al., 2014). However, previous works have limited understanding of how groups schedule and make decisions for events, or the influence of the group host and participants. Our system allows users to easily express their preferences by suggesting and voting for meeting locations and time. By analyzing group event planning process using data collected from our system, we can gain some useful insights for better group event organization.
3. System Design
In this section, we describe the design, architecture, and implementation of the OutWithFriendz system. We also present a walk-through example to illustrate the user workflow of our app for group event scheduling the typical user workflow through a series of use cases.
3.1. System Architecture
In order to understand group user behavior at scale, we designed versions of the OutWithFriendz mobile client for both the iPhone and Android platforms, allowing us to collect data on the scale of hundreds of users and user-generated invitations. This mobile system provides opportunity for us to track group event organization behavior on a regular, ongoing basis. We support the two most popular mobile platforms, iPhone and Android, allowing us to accommodate groups composed of users of both platforms. Our system consists of three major components: the mobile client, the data collection server, and the Google Cloud Messaging (GCM) server (Developers, 2014). We also call the Google Maps API to retrieve location search results. Figure 2 shows the overall architecture of our OutWithFriendz system.
The implementation of the OutWithFriendz system requires careful engineering to handle the communication and data synchronization between clients. Our OutWithFriendz server is implemented as a Java Web application using the Spring application framework (spring, 2017). All required functionality to the client is exposed through the server’s REST APIs. MongoDB is also used to store and manage all data on the server (MongoDB, 2017). To push notifications between server and clients, GCM services are used to handle all aspects of queueing of messages and delivery to client applications running on both mobile platforms. In addition, information about each location is obtained from Google Map Services (google, 2016), including the name, street address, and latitude/longitude coordinates.
3.2. UI Design Challenges
In order to enable a natural group decision-making workflow, we continuously streamlined the UI and workflow of the OutWithFriends app based on feedback collected from user studies. We started with an initial usage survey before releasing the app to the market. During our survey, we hired seven students on campus who have different academic backgrounds. They formed three groups to use our app and provided useful feedbacks for improving UI design. For example, these users suggested: (1) Adding a chat board to allow group members to discuss their opinions; (2) Allowing users to edit the location title and provide detailed information for each location; (3) Allowing users to link suggested locations with the Google Places application; (4) Pushing notifications if an invitation is created or modified; and (5) Replacing text buttons with interactive icon buttons. Implementing this functionality helped us improve our app to better support the real-life group event scheduling decision making process. We also added and altered application functionality to improve usability, based on many user suggestions received during application usage. At the beginning of the study, we focused on dining events only. Later we came to realize that the users would also like to use the OutWithFriends app for generic group gatherings, such as going for a hike or watching a movie. To support this functionality, we shifted from integrating with Foursquare API to the more suitable Google Places API. We also changed the workflow for the voting process to make it more flexible. Initially, invitation participants were required to decide on the meeting time before starting the voting process for the location. However, our users preferred to perform time voting and location voting concurrently, which is more flexible. Next, we describe the main workflow of our app.
3.3. A Walk-through Example Use Cases
To better understand the workflow of our mobile application, we provide a walk-through example of how the main functions are used for group event scheduling.
3.3.1. A host invites two friends to meet for dinner
In this use case, we describe the actions a user would take to invite some friends to meet for dinner. Here we call this user the host. When creating a new event invitation, the host will go to the window shown in Figure 2(a) and perform the following steps: (1) create a title for the invitation, such as Friday Dinner; (2) specify one or more possible dates and times for the invitation; (3) suggest meeting locations using Google Map Services, as shown in Figure 2(b); and (4) add one or more friends that want to be included as participants in the invitation. Finally, when the host is satisfied with the invitation settings, she taps the “send invitation” icon, to send the invitation to all selected participants. She can also start voting for her own preferences right after the new invitation shows up on her screen.
3.3.2. A user receives an invitation to meet several friends for dinner
First, the user receives a notification from the OutWithFriendz application indicating that she has received a new invitation. The user can express her preferences by voting on one or more possible options for meeting dates and locations, as shown in Figure 2(c). One important feature of our system is that the user may also add new proposed dates/time or locations to the invitation. Once the user has added a new option, it will be automatically made visible to all other participants. Users are also allowed to change their suggestions and votes throughout the voting process. In the “Chat” tab shown in Figure 3(a), the user is also able to send text messages to other group members for discussion and better coordination of the scheduling process.
3.3.3. Host finalizes the invitation based on voting results
The voting process continues until the host decides to finalize the meeting time and location. Only the host is permitted to finalize, which is shown in Figure 3(b). After the host has finalized the invitation, each participant receives a notification regarding this action. To support unforeseen changes, the host could still update the final decision after it is finalized. Each user’s main screen will show a list of invitations that she has participated in, as shown in Figure 3(c).
4. Data Collection and General Characteristics
We first describe the dataset we collected using our OutWithFriendz system, then conduct a data distribution analysis to understand the key characteristics of our dataset.
4.1. Data Collection
We deployed our OutWithFriendz mobile application on the Google Play and Apple Store marketplaces. To collect enough data for group dynamics analysis, we posted advertisements on Microworkers (Microworkers, 2017) and Craigslist (Craigslist, 2017) for participants. For teaching these users how to use our app correctly, we also made an introductory video, which is included in our supplemental file: “OutWithFriendzIntroductionVideo.mp4”. For each legitimate completed invitation, we paid the host of a group 20 dollars, with the provisions that: (1) The host and participants must live in the US; (2) The host should invite at least two other friends to the invitation using our app; (3) The group must demonstrate a full voting process; (4) The host must finalize the meeting time and location for the invitation; (5) Each participant would open their location services on their smartphone during the study and allow us to track their mobility traces; (6) At least half of the group members attended the finalized event111We added this requirement to prevent workers from creating fake invitations and making dishonest money. It would be interesting to analyze the low-attendance events, which we plan to investigate when we have more users and events.. From these two job post websites, we collected 246 legitimate invitations over a 5-month period from 432 users. In addition, 71 students on our campus used the app without getting any payment, which contributed another 76 legitimate invitations. The whole data collection period spanned from January 2016 to May 2017. In total, 503 distinct users of our OutWithFriendz application were identified, generating 322 legitimate invitations. 141 users have been the host of at least one event, and 72 hosts have created exactly one event each. Please note that each user is allowed to create and join multiple invitations in this study. Moreover, 11 groups (7 from paid users and 4 from students in our university) have used our mobile app frequently, with more than six legitimate invitations in our dataset. Figure 5 shows the distribution of all suggested locations recorded in our server across the US. It indicates that our users are widespread in 34 different states and 81 cities throughout the country.
To better understand the demographics of our users, we have conducted an anonymized survey using Google Form. We contacted each user through his/her Facebook Page to complete the survey. In total, 294 of our users have completed the survey. The results are shown in Figure 6. We can see that 60.8% of the users who responded to the survey are female and 48.5% are self-employed. Students, including some from our campus and some from the Microworker website, account for 27.5% of our participants. In addition, most (83.0%) of our users are young, aged 18-35.
4.2. Location Trace Data
In addition to collecting data about the event organization process, we also collected user mobility-related data. The OutWithFriendz app posts GPS user location traces to the server either every 5 minutes if the app is running in the background or every 30 seconds if the app is running in the foreground. Before participation all of our users were required to provide informed IRB consent. They would turn on the location services on their smartphone during the test so we were able to collect the data. All location traces are anonymized and permission to use this anonymized data is provided when installing the app. The total amount of user traces data collected was about 1.1 GB.
4.3. Group Size Distribution
left summaries the distribution of group size in our OutWithFriendz dataset. Here we define group size as the number of participants who finally stayed in the invitation. We do not count users who were removed from the invitation, either by themselves or by the host, because they did not participate in the whole scheduling process and their votes were not shown after the removal. We observe that most of the groups in our study are small, with the large majority being groups of three. We were pleased to see a significant fraction of groups with five (11.5%) and six members (4.2%) who were able to use the app concurrently. Our work focused primarily on obtaining data for groups of three or more, which we feel represent many typical social group interactions of interest to us. As a result, we did not focus on examining pairwise groups in this study. The figure’s trend lines suggest that if we had opened up our study to pairwise groups, then our data would have been overwhelmingly skewed toward pairwise groups. However, now that we have obtained substantial initial data for larger groups, we plan to also explore the behavior of pairwise groups in our future works.
4.4. Distribution of Days to Make Final Decision
We are interested in the duration that it took for event organizers to make their final decisions. As shown by the right figure in Figure 7, the number of days to make the final decision is somewhat evenly distributed, and there is no dominant duration in this distribution. This is a bit surprising or counter-intuitive, since we expected that there may be a more pronounced duration of decision-making within the first couple of days of creating an invitation. It is a little surprising that we find the distribution of days to make the final decision to be somewhat evenly distributed, which is illustrated in Figure 7 right. No duration dominates this distribution, while we expected that there might be a more pronounced bunching of decision-making in the first couple of days. However, there are also a substantial fraction of events that took four or more days to decide (about 30%), indicating that a large fraction of hosts are taking a long time to decide. This may be affected by the type of events and the amount of lead time. For daily meals, users can make a decision within thirty minutes while for some weekend activities, they will start planning it at the beginning of the week.
4.5. Voting Distribution of Individual Users
Voting distribution is based on the number of votes made per individual user. The distributions for time and place voting are shown in Figure 8. The majority of users will vote for one option as far as the event time. Similarly, the majority of users will vote for one option in terms of the place voting. In both processes, around 10% didn’t vote and 10% voted for more than 2 options. This voting behavior is analyzed further in later sections.
4.6. Distribution of the Proportion of the Votes
We also analyze the proportion of users who voted for event time or location in the final decision. We conduct an analysis that refers to the proportion of votes the poll winner achieved versus the other contenders. The distribution is shown in Figure 9. More than 70% of the final decisions winners for both time and location received majority votes to become the final choice. This is understandable since groups tend to agree on the majority votes. For the remaining 30%, we observe some very interesting behavior. In these polls, the final decisionswinners did not receive the majority of the votes. In fact, in a small fraction of cases, there is a non-zero proportion of polls in which the final decisionwinner received no votes. In these cases, the group host, who is the only one with the power to finalize the event time and location place/date, decided to override the majority voting results, either by personal fiat or possibly through a discussion with other group members that caused them to change their minds.
4.7. Suggestion Distribution
OutWithFriendz app allows group participants not only to vote for their preferences, but also to suggest new options. Figure 10 shows the suggestion distributions for host and participants. Most hosts will suggest 2 or more options for the event. We also observe a small portion of hosts who provide no options of their own, and rely on other group members to provide suggestions. For participants, more than 60% didn’t make new suggestions. They just vote for the existing options. Some made one new suggestion while very few of them would make too many new suggestions. We will further compare the influence of group host and participants in our group decision section.
We calculated the gender distribution of our OutWithFriendz users, which were about 42% male and 58% female. Please note that our app requires a user’s Facebook account to log in, but it is not permitted by Facebook to get the accurate gender information using Facebook user ID. So here we estimate users’ gender based on manually reading their Facebook name and profiles. It is possible that some users’ genders are miscounted here but we expect the proportion to be rare.
4.8. Metro vs Non-metro Areas
Using the location trace data we collected from our users and the U.S. census data, we are able to identify locations frequently visited by our users. The technique we used will be introduced in detail in Section 5. Then we can project each user’s home county using the frequently visited locations. According to the latest Rural-Urban Continuum Codes released in May 2013 (rur, 2013)
, every county is classified as a non-metro or metro area. In our dataset, 48% the users live in metro areas and 52% live in non-metro areas.222Please note our dataset contains 71 students who lived in Boulder doing this study. If we remove this student population, the proportion of metro and non-metro users would be 39% and 61%, all from crowdsourcing market users.
4.9. Weather Factor
An important external factor that can influence event organization relates to weather. Here we examine the impact of rain and temperature on our dataset. For this analysis, weather and temperature information for each event was scraped from weathersource API (weathersource, 2017) at its location and starting time in our dataset. Note that we can only get hourly weather data from weathersource. If the starting time of one event is 19:35, in the analysis we use 19:00 weather data at the same day crawled from weathersource. Here we decide it is raining if the precipitation of an event’s starting time is above 0. On snowy days, usually the precipitation will also be above 0, which we classify as rainy in our analysis.
Figure 11 shows the distribution of events that happened in rainy weather or not. 82.3% of events organized in our app occur in non-rainy weather. There are two reasonable explanations: (1) In many places around the country non-rainy days happen more often than rainy days; (2) Bad weather would have negative influence on real event attendance. Looking deeper, bad weather appears to affect the types of events that are organized. In OutWithFriendz, any place can be added to the Google Map as an option for voting, and need not be confined to a restaurant only. For example, people have used the app to organize events such as outdoor hiking and going to the movies. We divide all events into two category types: meal events and other events. Meal events refer to people hanging out for lunch or dinner, which is the majority event type in our dataset. Other events include activities that are not primarily dining, e.g. sporting and entertainment events. In our study, we found that bad weather would have less impact on meal events compare with other types of events. Figure 11 show that 66.7% of the events belong to meal events on non-rainy weather while this number goes up to 81.1% on rainy days.
5. Group Decision Analysis
The analysis in this section examines the impact of a number of factors on group decision. First, we define some concepts and notations that will be used throughout this section. In the following analysis, we only use completed invitations in our dataset. A “group decision” refers to the information submitted by the host after an event has occurred, including the final group consensus rating.
In addition, for each event , we define as the set of suggested meeting times and as the set of suggested locations. For each participant and option of event , we let be an indicator function which indicates whether voted for option :
Then we define user’s available time and location options as:
where and refers to user ’s time availability and location availability for event .
5.1. Impact of User Mobility on Group Decision
Using an individual user’s location trace data, we are able to analyze statistical properties of individual mobility. One way to consider a movement is to calculate the distance between two consecutive location trace points in our dataset. This will result in the detection of many very short movements, such as from one office to another in the same building. However, due to the location services limitation in today’s mobile phones, these short movements cannot be traced precisely. Figure 12 shows two examples of user traces recorded in our dataset. There are natural clusters in these location traces which appear to correspond to locations frequently visited by users, such as work, school, and home. To eliminate these very short movements and extract long movements, we implement an algorithm introduced by Ye et al (Ye et al., 2009), which was originally designed for GPS data. Assume that each individual’s location trace points detected by mobiles devices are ordered by timestamp . We identify two types of movements. Type 1 refers to the short movements of a user within a building. In Type 2, the user will travel from one area to another with a significant travel distance larger then , for some period of time. In our experiments, is set to 0.12 miles (200 meters) and the period threshold is set to 30 minutes, as suggested by (Ye et al., 2009). To extract all the Type 2 movements and eliminate Type 1 movements, we iteratively seek spatial regions where the user remains for more than 30 minutes and all the tracked points within this spatial region lie within 0.12 miles. Then the location points in this spatial region are fused together by calculating the centroid of these points. The centroid point is considered as a stationary point for the spatial region.
When the spatial regions are detected, we calculate movement distance as the travel distance from one region to another. We ignore all the trivial movements that occur within the same spatial region. Figure LABEL:fig:moves shows the distribution of single movement distances calculated by our proposed algorithm. The average movement distance is 5.35 miles on weekends and 4.12 miles on weekdays. This pattern seems reasonable, as people will be more likely to travel long distances for activities on weekends. The distribution of number of movements per day is shown in Figure LABEL:fig:trips. The average movement count on weekdays is 5.3 and 5.6 during weekends. We also observe that the proportion of very low movements and very high movements is higher on weekends than weekdays. One possible explanation could be that people prefer either rest on weekends, or to go out for activities on weekends, which results in either low-movement or high-movement days. We also compare our results with a US National Household Travel Survey conducted in 2009 (santos2011summary), which reports 4.2 movements during weekdays and 3.9 during weekends. Our numbers seem to be higher than the study in this survey. Our hypothesis is that some of our OutWithFriendz users may close their location services when they are not planning to use the app, in order to prolong batter life. These days are more likely to be their inactive days. Taking this into consideration, it is not surprising that the average movement detected by our app is slightly higher than reported by the survey.
We now examine the impact of user mobility on group behavior in OutWithFriendz. Here we define user mobility as the total travel distance traveled by a user in the 48-hour period preceding an invitation. Our assumption before was that users who traveled longer distances will be more exhausted, and thus less likely to have significant voting availability. However, our analysis refutes this conjecture:
Observation 1 ().
Users with higher mobility are more active in attending social events.
We use the Pearson correlation coefficient (Lawrence and Lin, 1989) to calculate the relationship between user mobility and voting availability. Table 1 shows that the correlation of user mobility with both date and location voting availability is positive, and the results are significant (). These results indicate that highly mobile users are more available for event attendance. There are two reasonable explanations for this phenomenon:
Previous studies have shown that users who travel by car, bus, and foot in daily life differ substantially in their value of time, in both revealed-preference and stated-preference surveys (Liu et al., 1997; Elgar and Bekhor, 2004). In our OutWithFriendz dataset, the users who travel long distances may travel by car. This increases their likelihood of attending events far away from their frequented spots.
Users who have higher mobility are more likely to be active event attenders. They are used to meeting with friends after school or work, which results in longer travel distances. Conversely, office workers who sit at their desks during the day have little mobility detected, but may still be tired after work and less likely to travel.
|Pearson correlation coefficient||p-value|
|The correlation of user mobility and date voting availability.||0.276||7.12e-05|
|The correlation of user mobility and location voting availability.||0.281||2.92e-06|
Observation 2 ().
Group mobility has a positive correlation with an area’s development degree.
Given the spatial regions that are detected, we are interested in investigating whether there exists any pattern between a group’s mobility and an area’s degree of development. Our hypothesis is that groups living in metro areas have higher mobility than groups living in non-metro areas, since metro group members may be more spread out in big cities and generate longer travel distances. To perform this analysis, we downloaded the 2016 U.S. area development degree data from the U.S. Census Bureau (Bureau, 2016). Here we use population density and number of housing units to calculate the development degree of an area. For simplicity, we consider the location of each group event and that area’s development degree. It is possible that group members live in a city but traveled to a rural area for the event. But this is rare in our dataset. Table 2 shows the relationship between group’s total travel distance and the corresponding county’s population density and housing units. The Pearson correlation coefficient for these two parameters are positive with p-values that are smaller than 0.05.
|Pearson Correlation Coefficient||p-value|
5.2. Impact of Individual Preference on Group Decisions
To discover underlying factors that may lead users to vote for specific event options, we first focus our analysis on individual users. A social event is typically characterized by two major factors: event time and location. Using the OutWithFriendz dataset we have collected, we first analyze the travel distance between event suggested locations and each participant’s closest location cluster, with the requirement that this cluster must contain a point with a timestamp that occurs within 2 hours before or after the finalized time for the invitation. The suggested location options are further divided into two categories: the location options with votes and location options without votes. Based on the results, we make the observation:
Observation 3 ().
Most users would like to vote for event locations near their frequented locations.
Figure 13 shows the cumulative distribution of travel distances among locations voted for and not voted for by each invitation participant. The average travel distance for voted locations is 4.19 miles while for non-voted locations is 7.53 miles. A Wilcoxon test found this to be a significant difference , which indicates users have clear preference to attend events near their frequented places. This is reasonable in daily life. For example, we would intuitively expect that users would prefer to go to dinner at restaurants that are close to their office or home.
Observation 4 ().
People like to attend social events after work on weekdays, while on weekends, events are distributed relatively evenly.
Additionally, we are also interested in investigating individual user’s temporal preference. Our hypothesis is that participants are more likely to attend events after work. Figure 14 depicts the suggested event times on weekdays and weekends. It is clear that in weekdays there is a high spike around 6pm. While in comparison, event times are distributed more evenly throughout the day on weekends.
5.3. Impact of Host Preference
In our OutWithFriendz system, the host has more authority than other participants. The host can not only decide who to invite, but also finalizes the event time and location. This suggests that the host will have more influence on the group decision-making process. In our dataset, we have several significant observations about host behavior.
Observation 5 ().
The final meeting location is closer to a host’s frequented place than other participants.
It’s not surprising that event host would show some “selfishness” when making the final decision. We calculated that the average distance between the final location and host’s closest frequented place is 5.23 miles. While the same metric for common participants is 6.75 miles, 29% longer than host, a significant difference according to a Wilcoxon test ().
|Final event date voted by host||0.71|
|Final event date voted by participant||0.36|
|Final event location voted by host||0.72|
|Final event location voted by participant||0.34|
Observation 6 ().
The probability that the final event date and location is voted by the host is significantly higher than that for other group participants. For events in which the host did not choose his/her voting option as the final decision, the main reason is to respect the majority voting results.
Table 3 shows the probability that final event option is voted for by the host and by another participant. It is clear that the final option is much more likely to have been voted for by the host than by other group members, with a probability of 0.71 vs 0.36 for the final event date (). and 0.72 vs 0.34 () for the final event location. We also observe that among all the invitations in which the final event time was not the host’s voting option, 95.2% coincided with the majority voting results. The percentage is 94.4% for the final event location. This indicates that, although hosts have a higher impact on making decisions, they still highly respect other group members’ opinions.
Observation 7 ().
The host choosing not to use the consensus voting result as the final decision would have negative influence on the event attendance rate.
In our OutWithFriendz application, the host can select a final decision that is contrary to the voting results. According to our user study, there are two main reasons for this behavior: (1) The option that received most votes is not suitable for the event host; (2) the users discussed through using the app’s chat function and some members changed their minds but did not update their votes. In our OutWithFriendz dataset, 7.3% of final dates and 9.2% of final locations are contrary to voting results. We calculated the Pearson correlation between whether the host complies with the consensus opinion and the corresponding event attendance rate. The results are shown in Table 4. The positive correlation is significant here for both location voting and date voting. These results confirm that for event organization, hosts that don’t comply with voting results have negative impact on the attraction of participants.
|Whether host comply location voting results and event attendance rate||0.48|
|Whether host comply date voting results and event attendance rate||0.47|
5.4. Voting Process Analysis
Voting is one of the most innovative aspects of our OutWithFriendz system. In contrast to traditional online event organization services, such as Meetup and Douban Events, where the meeting location and time is decided only by group host when the invitation is created, OutWithFriendz allows all group members to express their preferences through suggestions and votes. After all invitees have responded to the poll, the group host is able to find a mutually agreeable location and time that usually accommodates most of the group members. Tracking the group’s voting process using our system offers a great opportunity to study group decision making behavior.
Observation 8 ().
Early voters tend to vote for a wide variety of options, while later voters are more likely to report limited availability.
Figure 15 shows the relationship between group members’ average availability and their “voting position”. A user’s availability is defined by Equations 2 and 3. Voter position refers to the temporal index of casting votes within the scope of an invitation. The host’s position is 1, the first voter’s position is 2, and so on. There is a clear decrease of availability as voter position increases, and this result is consistent for both location voting and time voting. There are several possible explanations for this observation:
People who came to the poll later may be busier than early voters, and had a smaller time window before the actual event time, thus their availability is more limited compared with early voters.
The polls in the OutWithFriendz application are all open polls, which means later voters can see the current voting results. Their votes may not be able to change the current status significantly because every voter can only vote once for a given option.
Late voters will vote only for agreeable options that help the host to more easily finalize decisions. This phenomenon will be further discussed by the next observation.
Observation 9 ().
Late voters tend to vote for options that align with existing voting results and are mutually agreeable.
Due to the fact that new voters can observe other voters’ responses, these early responses would easily affect future voting behaviors. In our dataset, we find that later voters are more willing to vote for options that coincide with existing voting results, which makes it easier for the host to find common mutually agreeable options. Here we define the voting coincidence by cosine similarity:
refers to the new voter’s voting vector, andrefers to the existing voting vector. For example, if there are four date options in invitation , and they receive 1, 3, 1, and 0 votes respectively, then the is . If a new voter votes for second, and fourth option, then is [0, 1, 0, 1]. The coincidence here is the cosine similarity between and , which is 0.640. Figure 16 shows the relationship between average voting coincidence and voter position. It is clear that there is a positive relationship between voter position and coincidence in both date voting and location voting. Later voters will try to consider their options in light of the whole group’s voting behavior. Sometimes, these later voters may vote for less convenient options in order to make the host’s life easier.
In this section, we summarize the key results of our analysis
and provide insights into how these results can be utilized for better
group event planning experiences. The results cover the impact of user mobility,
individual preference, and the host preference on the group event scheduling process.
We also discuss the impact of user behavior during the voting process.
User Mobility. The analysis in Section 5.1 showed that users with higher mobility are more likely to be active participants in group events. They are more active in voting for proposed event location and event time. We attribute this behavior to the capability of these users to travel greater distances (perhaps due to owning a car or having more available time), which gave them flexibility in selecting location and time options for events. One way to use this observation for better event planning would be to recommend more diverse locations and dates for highly mobile users, as they tend to be more willing to explore new options. On the other hand, users with less mobility should not be overwhelmed with a large number of choices. In addition to recommending event locations and time, this observation can be used for forming groups by matching users with similar mobility levels in the same group, which can lead to smoother event planning experiences.
Individual Preference. The analysis in Section 5.2 revealed typical patterns related to individual preferences for event time and locations. First, with regard to event location, users tend to arrange events at nearby locations to avoid traveling long distances. Second, with regard to event time, on weekdays, users want to schedule events after their working hours, while on weekends users show more flexibility. These observations are worth considering for smarter group event planning. For event locations, the application could suggest places such that the mean travel distance for the group members is minimized, so as to provide a reasonable compromise for the whole group. Likewise, suggested event time should occur outside of typical working hours of the group members.
Host Preference. The results in Section 5.3 show that group event planning is heavily influenced by the host who creates the invitation. From the analysis in this section, it is evident that the final event location is on average closer to the host’s frequented locations than that of other group members. We also see that the final event locations and dates are more likely to be the options voted by the host. However, the influence of the host can also lead to negative outcomes: when the host chooses not to follow the group’s consensus, event attendance is reduced. These effects point to the need to carefully consider the preferences of the host, and how these preferences align with the preferences of the group, when providing recommendations to event participants. Effective communication mechanisms between the host and the participants should also be provided.
Voting Analysis. The analysis in Section 5.4 shows that the votes cast by early voters are very likely to affect late voters. Late voters tend to vote for fewer options, and these options tend to match those that have already received votes from early voters. This phenomenon can be used to improve the event planning experience. For example, we could encourage users to vote early, so that their votes will carry more weight. We could also hide existing voting results, so as to prevent existing votes from biasing later voters, and facilitate the voting process by providing voting recommendations to users based on their historical voting patterns.
Expanding event coverage. Though we have collected more than 300 completed group events, we hope to grow our user base through more effective advertising, so that we may achieve viral adoption and gather data at even larger scales.
Facebook friends requirement. To obtain a user’s friend list, OutWithFriendz currently only allows users to log in through their Facebook accounts. Users may also want to invite people who are not already a Facebook friend or do not use Facebook at all. This limits our app’s ability to support larger groups. We plan to design an “Add Friend’ function which enables users to log in and connect with other users directly within the application.
Different voting mechanisms. All the polls in OutWithFriends are currently designed to be open polls, which means late-coming voters can see existing voting results, which may influence their vote. In a future version of the application, we plan to add a closed poll option. For closed polls, existing voting results will be hidden from new voters. This functionality will allow us to examine how a closed poll mechanism influences the group decision-making process.
Group recommendation. We have seen from our work on OutWithFriendz that a number of factors influence group decision making behavior. We believe that group context could be seen as inhabiting a latent trait space, similar to how users inhabit a latent user trait space in the matrix factorization framework for individual recommendation. Furthermore, our work has revealed that host and individual influence within a group plays a significant role in the group decision-making process. In future work, we intend to pursue the development of a group recommendation system that incorporates these ideas into a probabilistic model for group preference behavior. Building a distributed system implementation of mobile group recommendation will help us gain a better understanding of mobile group dynamics in the real world, and provide useful suggestions for group organizers.
7. Conclusion and Future Work
In this paper, we have presented the results of a large-scale study of OutWithFriendz, a newly designed mobile application for group event scheduling organization. We summarize our key findings as follows: (1) User mobility has a significant impact on group event attendance. (2) Users would like to vote for locations near their frequented places. On weekdays, they would like to meet after work while on weekends, they have wider time options. (3) A group host has a higher impact on the group decision making process than other members. (4) Early voters are more likely to vote for more options while late voters tend to coincide with existing results to find a mutually agreeable option. We believe the results presented in this paper are a good start towards better accurate understanding of group event scheduling behaviors in real life. Our analysis of this first-of-a-kind real world user study of mobile group dynamics has provided very important guidance for group event scheduling and recommendation mobile group recommendation and mobile group formation.
We plan to pursue several directions for future work. First, although we have collected more than 300 completed group events, we hope to grow our user base through more effective advertising, so that we may achieve viral adoption and gather data at even larger scales. Second, to obtain a user’s friend list, OutWithFriendz currently only allows users to log in through their Facebook accounts. Users may want to invite people who are not already a Facebook friend or do not use Facebook at all. This limits our app’s ability to support larger groups. We plan to design an “Add Friend” function which enables users to log in and connect with other users directly within the application. Third, currently, polls in OutWithFriends are designed to be open polls, allowing late-coming voters to see existing voting results, which may influence their own votes. We plan to add a closed poll option. For closed polls, existing voting results will be hidden from new voters. This functionality will allow us to examine how a closed poll mechanism influences the group event scheduling process. Lastly, we have seen from our work on OutWithFriendz that a number of factors influence group decision making. We believe that group context can be seen as inhabiting a latent trait space, similar to how users inhabit a latent user trait space in the matrix factorization framework for individual recommendation. Furthermore, our work has revealed that both host and individual members within a group play an important role in the group event scheduling process. In our future work, we intend to pursue the development of a group recommendation system that incorporates these ideas into a probabilistic model for group preferences and make group event recommendations in real-world settings. This will help us gain a better understanding of group event dynamics and provide useful suggestions for group event organizers.
- rur (2013) 2013. Rural-Urban Continuum Codest. https://www.ers.usda.gov/data-products/rural-urban-continuum-codes/. (2013).
- Beckmann and Gross (2011) Christoph Beckmann and Tom Gross. 2011. AGReMo: providing ad-hoc groups with on-demand recommendations on mobile devices. In Proceedings of the 29th Annual European Conference on Cognitive Ergonomics. ACM, 179–182.
- Brown et al. (2014) Chloë Brown, Neal Lathia, Cecilia Mascolo, Anastasios Noulas, and Vincent Blondel. 2014. Group colocation behavior in technological social networks. PloS one 9, 8 (2014), e105816.
- Bureau (2016) US Census Bureau. 2016. US Population Density Dataset 2010-2016. (2016).
- Craigslist (2017) Craigslist 2017. Craigslist. https://www.craigslist.org/. (2017).
- Developers (2014) Android Developers. 2014. Google cloud messaging for Android. (2014).
- Doodle (2017) Doodle 2017. Doodle. http://doodle.com/. (2017).
- Du et al. (2014) Rong Du, Zhiwen Yu, Tao Mei, Zhitao Wang, Zhu Wang, and Bin Guo. 2014. Predicting activity attendance in event-based social networks: Content, context and social influence. In Proceedings of the 2014 ACM International Joint Conference on Pervasive and Ubiquitous Computing. ACM, 425–434.
- Elgar and Bekhor (2004) Alon Elgar and Shlomo Bekhor. 2004. Car-rider segmentation according to riding status and investment in car mobility. Transportation Research Record: Journal of the Transportation Research Board 1894 (2004), 109–116.
- Evite (2017) Evite 2017. Evite. https://www.evite.com/. (2017).
- facebook (2017) facebook 2017. Facebook Events. http://events.fb.com/. (2017).
- Fischer et al. (2013) Joel E Fischer, Stuart Reeves, Stuart Moran, Chris Greenhalgh, Steve Benford, and Stefan Rennick-Egglestone. 2013. Understanding mobile notification management in collocated groups. In ECSCW 2013: Proceedings of the 13th European Conference on Computer Supported Cooperative Work, 21-25 September 2013, Paphos, Cyprus. Springer, 21–44.
- google (2016) google 2016. Google Maps Services. https://developers.google.com/maps/. (2016).
- Goyal et al. (2010) Amit Goyal, Francesco Bonchi, and Laks VS Lakshmanan. 2010. Learning influence probabilities in social networks. In Proceedings of the third ACM international conference on Web Search and Data Mining. ACM, 241–250.
- Goyal et al. (2011) Amit Goyal, Francesco Bonchi, and Laks VS Lakshmanan. 2011. A data-based approach to social influence maximization. Proceedings of the VLDB Endowment 5, 1 (2011), 73–84.
- Jayarajah et al. (2015) Kasthuri Jayarajah, Youngki Lee, Archan Misra, and Rajesh Krishna Balan. 2015. Need accurate user behaviour?: pay attention to groups!. In Proceedings of the 2015 ACM International Joint Conference on Pervasive and Ubiquitous Computing. ACM, 855–866.
- Lampinen et al. (2009) Airi Lampinen, Sakari Tamminen, and Antti Oulasvirta. 2009. All my people right here, right now: Management of group co-presence on a social networking site. In Proceedings of the ACM 2009 international conference on Supporting group work. ACM, 281–290.
- Lawrence and Lin (1989) I Lawrence and Kuei Lin. 1989. A concordance correlation coefficient to evaluate reproducibility. Biometrics (1989), 255–268.
- Li et al. (2013) Yanhua Li, Wei Chen, Yajun Wang, and Zhi-Li Zhang. 2013. Influence diffusion dynamics and influence maximization in social networks with friend and foe relationships. In Proceedings of the sixth ACM international conference on Web Search and Data Mining. ACM, 657–666.
- Liu et al. (1997) Rongfang Liu, Ram Pendyala, and Steven Polzin. 1997. Assessment of intermodal transfer penalties using stated preference data. Transportation Research Record: Journal of the Transportation Research Board 1607 (1997), 74–80.
- Lucero et al. (2015) Andrés Lucero, Danielle Wilde, Simon Robinson, Joel E Fischer, James Clawson, and Oscar Tomico. 2015. Mobile collocated interactions with wearables. In Proceedings of the 17th International Conference on Human-Computer Interaction with Mobile Devices and Services Adjunct. ACM, 1138–1141.
- meetup (2017) meetup 2017. Meetup About. http://www.meetup.com/about/. (2017).
- Microworkers (2017) Microworkers 2017. Microworkers. https://ttv.microworkers.com/. (2017).
- MongoDB (2017) MongoDB 2017. MongoDB. https://www.mongodb.com/. (2017).
- Park et al. (2008) Moon-Hee Park, Han-Saem Park, and Sung-Bae Cho. 2008. Restaurant recommendation for group of people in mobile environments using probabilistic multi-criteria decision making. In Asia-Pacific Conference on Computer Human Interaction. Springer, 114–122.
- Pramanik et al. (2016) Soumajit Pramanik, Midhun Gundapuneni, Sayan Pathak, and Bivas Mitra. 2016. Predicting Group Success in Meetup. In Tenth International AAAI Conference on Web and Social Media. 636–666.
- Sen et al. (2014) Rijurekha Sen, Youngki Lee, Kasthuri Jayarajah, Archan Misra, and Rajesh Krishna Balan. 2014. Grumon: Fast and accurate group monitoring for heterogeneous urban spaces. In Proceedings of the 12th ACM Conference on Embedded Network Sensor Systems. ACM, 46–60.
- spring (2017) spring 2017. Spring Framework. https://projects.spring.io/spring-framework/. (2017).
- weathersource (2017) weathersource 2017. Weather Source API. https://weathersource.com/products/onpoint-api/. (2017).
- Ye et al. (2009) Yang Ye, Yu Zheng, Yukun Chen, Jianhua Feng, and Xing Xie. 2009. Mining individual life pattern based on location history. In Mobile Data Management: Systems, Services and Middleware, 2009. MDM’09. Tenth International Conference on. IEEE, 1–10.
- Yu et al. (2015) Zhiwen Yu, Rong Du, Bin Guo, Huang Xu, Tao Gu, Zhu Wang, and Daqing Zhang. 2015. Who should i invite for my party?: Combining user preference and influence maximization for social events. In Proceedings of the 2015 ACM International Joint Conference on Pervasive and Ubiquitous Computing. ACM, 879–883.
- Zhang and Lv (2017) Shuo Zhang and Qin Lv. 2017. Event Organization 101: Understanding Latent Factors of Event Popularity. In Eleventh International AAAI Conference on Web and Social Media.
- Zou et al. (2015) James Zou, Reshef Meir, and David Parkes. 2015. Strategic voting behavior in doodle polls. In Proceedings of the 18th ACM Conference on Computer Supported Cooperative Work & Social Computing. ACM, 464–472.