Human factors, such as emotions, personality traits and trust network, have been proved to play an important role in the decision making process. The impact by personality in individual and group decision making is still under investigation, especially in the area of educational learning. In this paper, we propose two approaches to distinguish the \dominator " and \follower" in group decision making by using an educational data. Our experiments also show that the characteristics of these two user roles can further be utilized in group recommender systems to produce better item recommendations.
Recently, the importance of personality is realized not
only for predicting the individual tastes, but also the
group preferences. For example, a group of users may
decide which dishes should be ordered for a group lunch.
Or, a group of tourists would like to make a decision
about the list of points of interests for tomorrow's trip.
Researchers [
HUMANIZE '18, March 11, 2018, Tokyo, Japan to group decisions, but some of other users (i.e., dominators) may play a dominant role in group decision making. The impact by personality in individual and group decision making is still under investigation, especially in the area of educational learning. For example, team work becomes more and more popular in the educational learning. Students may be suggested to work together on the assignments or projects. Decision making is involved in such a scenario, e.g., how will the students build the team, or which materials or topics should a team select to start learning, etc. Furthermore, it is also interesting to understand which group of the users may yield to group decisions. In this paper, we discuss our analysis to distinguish followers and dominators in the group decision by using an educational data. Our contributions can be summarized as follows:
We propose two approaches to identify the dominators and followers in the group decision making.
We discover and summarize the characteristics of these user roles in terms of the personality traits.
RELATED WORK
Personality has been successfully applied to improve
decision making in di erent areas, such as tourism [
ITMLEARNING PLATFORM
The impact of personality on individual and group
decisions is under investigation in the area of educational
learning. But unfortunately, there are no available
data sets for public research in this domain. Even in the
general area of group recommendations, most of the
research may use the MovieLens data { the evaluation is
usually based on the simulated groups. In this case, we
start collecting our own data for the research purpose.
ITMLearning platform is built for the department of
information technology and management (ITM) at the
Illinois Institute of Technology in USA. The platform is a
technology-enhanced learning system which aims to: a)
suggesting appropriate learning materials (e.g., books,
articles, tutorials, videos); b) recommending job
positions; c) assisting instructors in the teaching.
One of the ongoing projects from this platform is
collecting students' preferences on the topics of the projects in
order to better support learning and assist teaching [
Topics of The Projects: First of all, we provide a list of potential topics for each course respectively. Take data analytics course for example, we provide the information about 50 data sets that are available on Kaggle.com. Students should select one of them, de ne the research problems, and gure out solutions by using the data analytics skills.
Collection of Individual Preferences: At the beginning, each student is required to ll the questionnaire by himself or herself. Each subject should select at least three liked and disliked topics of the projects, and provide an overall rating to them. In addition, they are asked to rate each selected project on three criteria: how interesting the application area is (i.e., App), how convenient the data processing will be (i.e., Data), how easy the whole project is (i.e., Ease). The rating scale is from 1 to 5.
student has to decide whether they will complete the
project individually. For the team work, they need
to nd partners and build the team by themselves.
Each team will ll the same questionnaire from the
perspective of a team based on the group discussions.
In addition to these preference data, we collect
demographic (e.g., age, gender, marriage status, home
country) information and personality traits of each
student. We choose the Big Five Factor (Big5) [
The full questionnaire includes the ten statements that are listed below, and the subjects are asked to give a rating in scale 1 (strongly disagree) to 7 (strongly agree) to each of them.
At this moment, we have collected data for a full year { we obtain a data set with 194 individuals and 122 groups. 81 out of 122 groups are composed of more than one members. More speci cally, 60% of these 81 groups are composed of two members, and the remaining groups are composed of three or four members. The individuals leave 1951 ratings on the topics of projects, while the groups leave 745 ratings in total. In addition to the overall ratings, we collect their ratings on three criteria as introduced above. For the purpose of personalization, this data is available for traditional recommender systems (i.e., recommendations for individuals), group recommender systems (i.e., recommendations for each group), and multi-criteria recommender systems (i.e., recommendations based on multi-criteria decision making), as well as context-aware recommendations (i.e., semester, year, course can be viewed as the context information). The project is still ongoing and we expect to collect more data gradually.
ANALYSIS AND DISCUSSIONS Personality Traits by Gender In our data, 42% of the subjects are female. We'd like to explore whether there is a signi cant di erence in their personality traits in comparison with males. Table 1 presents the mean and standard deviation (SD) of the scores in the Big5 factors for the overall, male and female individuals respectively. In addition, we also observe that the standard deviations in neuroticism and extraversion are signi cantly larger than other personality factors. The two-independent sample hypothesis tests reveal that the di erence on conscientiousness and agreeableness between males and females are signi cant at the 95% con dence level. Team Building We further analyze the 81 groups which are composed of at least two members. Students actually nd their own partners and build the team without intervention by the instructors. We are pretty interested in how they build a team or what are the most important criteria for them to select partners. More speci cally, we try to measure the intra-group similarities.
First of all, each subject can be represented by the Big5 vector. Cosine similarity, as shown by Equation 1, can be used to produce the similarity between two subjects Ua and Ub in a same team. The vectors V!a and !Vb are the Big5 vectors for Ua and Ub respectively. We obtain similarity values of each pair of the subjects in a team, and the mean similarity is viewed as the intra-group similarity.
Sim(Ua; Ub) = ! !
Va Vb ! jjVajj2
! jjVbjj2 (1)
Furthermore, each subject is required to provide the individual preferences (i.e., user ratings) on the topics of the projects. Alternatively, we can represent each subject by his or her rating vector. The rating vector can be lled by the overall rating or the multi-criteria ratings on app, data and ease respectively. In other words, V!a and !Vb could be rating vectors based on the overall rating or the multi-criteria ratings. The similarity between two subjects can be obtained by the Equation 1 accordingly. As a result, we are able to produce the intra-group similarities by representing a user as the Big5 vector or a rating vector. We further analyze the distribution of these intra-group similarities, and visualize them as box plots in Figure 1. It is clear that the intra-group similarity is signi cantly higher by the representations based on the Big5 factors than the ones based on user's rating vectors. It implies that the subjects prefer to nd the team members by the personality traits, even if their tastes on the projects may be di erent. The average intra-group similarity based on the rating vectors is actually below 0.5, which is surprising.
Distinguish Dominators and Followers It has been recognized that personality can a ect group decisions. Our goal is to nd out distinct individuals who react di erently in group decisions. More speci cally, we de ne dominator(s) as one or more members in a group who are the decision leaders, and follower(s) as the member who may yield to group decisions. We try to these two user roles from the perspective of usergroup similarities and user-group con icts which can be further discussed as follows. Also note that our following analysis is based on the 81 groups which is composed of at least two team members.
We have both individual and group preferences on the topics of the projects. Each individual and group can be represented by the rating vectors. In this analysis, we focus on the overall rating only and ignore the multicriteria ratings for simplicity. The similarity between a group and an individual in the group can be computed by the cosine similarity of the representations based on the Agreeableness*
Extraversion rating vectors. If this similarity value is relatively low, it implies that this subject yields to the group decisions. Subjects with higher user-group similarity can be viewed as the \dominator", while the subjects with user-group similarity smaller than a threshold can be the \follower".
Or, we will set two thresholds. For example, we obtain the 1st and 3rd quartile of the user-group similarities. The subjects with user-group similarity larger than 3rd quartile will be viewed as dominators, while the users with user-group similarity smaller than 1st quartile will be considered as the followers.
We found that the second way was better, therefore we only present these results in the following sections. Afterwards, we computer the mean Big5 vector for the subjects as dominators and followers which can be depicted by the radar chart as shown in Figure 2. We use \Global" to represent the mean Big5 vector of all the subjects. The \*" denotes a signi cant di erence (two-independent sample test at 95% con dence level) in a speci c personality trait between dominators and followers. We can observe that the signi cant di erence only shows up in agreeableness, while dominators actually have larger degree of agreeableness. It sounds surprising to us, since we expect the followers may yield to the group decisions and they should present relative larger degree of agreeableness.
After a further investigation, we realize that the cosine
similarity based on the rating vectors relies on the
number of co-ratings by a team and an individual in the team
{ the similarity may be not reliable if the number of
corated items is limited. In our data, the average value of
co-ratings by the teams and the team members is 3.33
with standard deviation 2.45. We believe the results in
Figure 2 are not reliable due to the limited number of
co-ratings between a team and team member.
We gure out a way to alleviate this problem. More
speci cally, we blend user ratings and group ratings
together, while each group is viewed as a special user. We
utilize a matrix factorization model based on this rating
matrix to nd the best model which can minimize the
squared prediction errors in the ratings. Finally, each
user and each group can be represented by a latent vector
which is learned by the matrix factorization model. The
user-group similarity, therefore, can be calculated by the
cosine similarity of two latent factors. In our work, we
use the biased matrix factorization [
Figure 3 presents new comparisons of the BIG5 traits. The dominators and followers are identi ed based on the same method as mentioned above, where we represent a team and a team member as the latent vectors learned based on the matrix factorization model. We can observe that there are signi cant di erences between dominators and followers in openness, agreeableness and extraversion based on the two-independent sample test at 95% con dence level. More speci cally, dominators present higher values in the openness and extraversion, while the agreeableness value is relatively higher in the followers who may yield to the group decision. It is not surprising to see that a dominator could be more extraverted since he or she may be a talkative, con dent and assertive person. In terms of the openness, one explanation could be that dominator is usually the rst person to start the discussions in a team, and they may produce novel ideas and lead the group decisions. By contrast, the followers present larger degree of the agreeableness, which may infer that they tend to accept the group decisions even if they have di erent opinions.
Alternatively, we can distinguish dominators and
followers based on the notion of \con icts". Recio-Garcia, et
al. [
More speci cally, we de ne con ict as either the false positive or the false negative case. A \false positive" case can be described as the situation that a subject presents positive preference on one item, but his or her group nally made a negative decision on the same item. Accordingly, the scenario that being positive on one item by group decision but negative by a team member will result in a \false negative" case. In our experiment, we use a rating threshold to de ne whether it is positive or negative. More speci cally, it is positive when individual or group rating on one project is larger than 3. We compute the total number of con icts (including both false positive and false negative cases) for each team member in a team. We nd out that only 22.5% of the subjects present the con icts in our data.
Accordingly, we obtain the mean, 1st and 3rd quartile of the number of con icts, and set the threshold to distinguish the followers and dominators. The process is similar to the one we used to identify di erent user roles by using the user-group similarities { we can use a single threshold or two thresholds. In our experiments, we nally use the mean value of the number of con icts as the threshold to split the subjects to dominators and followers.
The sparsity problem is involved again since the number of co-ratings by a team and a team member is limited. We use matrix factorization model to make predictions on the unknown ratings for both subjects and teams. As a result, 97% of the subjects present the con icting behaviors. We use the single threshold to separate the subjects to dominators and followers, while the comparison in BIG5 can be depicted by Figure 4. We can observe that the statistical signi cance only presents in the agreeableness, while the followers usually have higher values in agreeableness.
We try to distinguish the dominators and followers in the group decision making by two proposed approaches { one is by the similarity between the team and its team members, another one is by the con icts between individual and group preferences. We nd that openness, agreeableness and extraversion are the three in uential factors to recognize the dominator and followers by using the user-group similarities. By contrast, agreeableness is the only crucial factor we nd by using the method based on the con icts.
Some previous research have also identi ed the
important personality traits in the group decisions. For
example, Ertac, et, al. [
Least misery (LM): It is used to minimize the misery for the group members. P (g; t) is the minimal predicted rating by the team members.
Most pleasure (MP): It tries to maximize the
happiness or pleasure for the group members. P (g; t) is the
maximal predicted rating by the team members.
For the purpose of rating predictions, we use the biased
matrix factorization [
The recommendation performance based on the MAE metric can be presented in Table above. We can observe that the AVG method is the best one among all of the baseline approaches. By incorporating the identi ed dominators and followers, the method by user-group similarities can o er signi cant improvement for the AVG and MP strategies. The method by con icts obtains improvements for the AVG method only. Note that the signi cance test was examined at the 90% con dence level. Unfortunately, there are no signi cant improvements at the 95% level, and the improvement is relatively small. Furthermore, the method by user-group similarities presents signi cant improvements in MP rather than LM. It implies that the followers may leave false positive contributions to the group decisions in our data. It is because the performance can be improved if we ignore the followers in the MP method.
CONCLUSIONS & FUTURE WORK In this paper, we try to distinguish the dominators and follows in group decision making by using user-group similarities and con icts. We further take advantage of the identi ed dominators and followers in di erent group recommendation strategies. And we nd that we are able to obtain signi cant improvements by ignoring the followers in the group recommendations. Furthermore, we nd that students with similar personality tend to work together in our data.
This paper presents our initial work, while there are
plenty work to do. For example, the identi cation method
by using user-group con icts relies on the user ratings.
However, user bias should be taken into account in the
process to de ne the false positive and false negative
cases. The approach by user-group similarities is dependent
with the recommendation algorithms, since we use the
algorithm to ll in the unknown ratings. We will further
explore the corresponding solutions in the future.
Furthermore, we evaluate the recommendation performance
based on simple group recommendation strategies, but
there are several advanced work which can directly
incorporate personality in the group recommenders, such
as the work by [