Abstract
Using digital tools for teaching allows to unburden teachers from organizational load and even provides qualitative improvements that are not achieved in traditional teaching. Algorithmically supported learning group formation aims at optimizing group composition so that each learner can achieve his or her maximum learning gain and learning groups stay stable and productive. Selecting and weighting relevant criteria for learning group formation is an interdisciplinary challenge. This contribution presents the status quo of algorithmic approaches and respective criteria for learning group formation. Based on this theoretical foundation, we describe an empirical study that investigated the influence of distributing two personality traits (conscientiousness and extraversion) either homogeneously or heterogeneously on subjective and objective measures of productivity, time investment, satisfaction, and performance. Results are compared to an earlier study that also included motivation and prior knowledge as criteria. We find both personality traits to enhance group satisfaction and performance when distributed heterogeneously.
1 Introduction
When teaching small classes, teachers with good knowledge of their students’ competencies and interests can form optimized learning groups for certain learning goals so that all students benefit the most from collaborative work. However, with increasing numbers of students or when knowledge about competencies is limited, manual group formation becomes a complex problem that needs computational assistance. Blended learning scenarios or scenarios of pure online learning render it impossible to conduct group formation without algorithmic support.
Common practice is using either randomized grouping or self-grouping by the students. Both approaches have major disadvantages [20] as they neglect the opportunity to optimize for the specific learning goal. From an algorithmic point of view there are several approaches addressing the problem of learning group formation. Some approaches make use of machine learning algorithms in order to optimize group formation on the basis of prior collaborative work. Other approaches are based on numerical optimization algorithms which use vectors of learners‘ criteria and appropriate boundary conditions.
Beside the question of the most suitable algorithm, another question is which criteria are relevant for group formation and which goals of optimization are affected by them. Criteria can be categorized by their stability. Instable criteria have a direct relation to a certain learning scenario and the tasks involved in that scenario. Examples are the degree of participation in a discussion or ratings of group partners after task completion. Such instable criteria can be measured relatively easily in digital systems (such as learning management systems, LMS, e. g., Moodle[1]). However, relevance for group performance in different scenarios or in different group constellations is limited due to the lack of transferability.
Stable criteria are person-related features (e. g., personality traits) that are not prone to alteration when learning scenarios or group composition change. Therefore, effects of stable criteria on quality of group work can be investigated repeatedly in different scenarios, thereby increasing the possibility to generalize the findings. Educational psychology has investigated which stable criteria can be used for group formation and the possible outcomes for quality of learning groups.
Research on personality traits largely builds on the concept of the “Big Five”: extraversion, conscientiousness, openness, social agreeableness, and neurotizism. These traits are considered to be relatively stable over the live span [19] and can be measured reliably by means of questionnaires [24]. In the context of group formation, these traits have been investigated more than any other personality traits, but still the findings are inconclusive towards the question whether one or more of these five traits should be distributed homogeneously or heterogeneously within learning groups. Researchers have proposed that two of the big five predominantly influence the success of learning groups: Extraversion (which is associated with leadership behavior) is supposed to be distributed heterogeneously in order to minimize conflicts. Conscientiousness (which is associated with goal setting) is supposed to be distributed homogeneously in order to facilitate agreement on a common group goal. However, these hypotheses are derived on the basis of correlational studies; experimental studies that allow for causal inferences are largely missing.
A pilot study with experimental design was conducted by the authors [16]. In this study, a combination of several criteria was applied in the experimental group, contrasted with a control group that was grouped randomly. The criteria in the experimental group included extraversion (heterogeneously) and conscientiousness (homogeneously), along with several other criteria that were all applied simultaneously. While this combination of criteria proved to be successful in improving group work, it was not possible to investigate the relative impact of the different criteria. In the present study, we therefore exclusively featured two criteria, namely extraversion and conscientiousness, and independently varied whether each of them was applied or not, and if so whether their distribution was homogeneous or heterogeneous. This elaborated 3×3 experimental design might guide future research on the impact of group formation interventions on measures of quality such as learning gains, satisfaction, and group performance.
In the present paper, we will provide an overview of insights from educational psychology concerning possible criteria for group formation (paragraph 2.1). We then discuss the algorithmic challenges of computer-supported group formation (paragraph 2.2), followed by a detailed presentation of the solution proposed in our software MoodlePeers (paragraph 2.3). Paragraph 2.4 presents a pilot study conducted with this software. Paragraph 3.1 will display the hypotheses we derived from the current state of the art, followed by the description of our research design and methods (paragraph 3.2). A concise look at the major findings of our study will be given in paragraph 4. Results are discussed with respect to implications for both future research and practical applications (paragraph 5).
2 Related Work in the Field of Learning Group Formation
To achieve the goal of optimized learning group formation, it is relevant to delineate this group formation problem from team formation in organizational contexts. Nevertheless, some of the related work algorithms are based on either solution for the so called “Cell Formation Problem” or on solutions from capacity planning. Common to these solutions is the optimization towards the results, i. e. the output of the work process while utilizing the available resources ideally in optimal combination. From a didactical point of view the optimization of final products is indeed not the inherent goal of learning group formation, but rather maximization of each individual leaner’s learning progress. Consequently, the optimal combination of differences and similarities of learners within a group, based on findings from pedagogical psychology about significant factors, is crucial. Furthermore, a second major didactical aspect is a preferably equal quality of the formed groups to provide to every learner the same chances for learning progress.
“Therefore, we define the learning group formation problem as the optimal formation of learning groups based on a given set of learners to be grouped, while considering manifold criteria – each with different criteria-related optimization goals, combined with simultaneous compliance to a[n] equalized quality of all formed groups”
(cf. [17, p. 17]).
2.1 Findings in Educational Psychology
Various empirical studies have shown cooperative learning to be a very effective method [18]. The advantages over individual learning concern not only subjective aspects such as satisfaction but also objective achievements. Less intensively researched in this context is the question of the criteria according to which the study group members should be selected. Correlative studies have shown that demographic characteristics such as gender, age or educational attainment have a relatively weak correlation with the success of groups [9]. In contrast, stronger relations were uncovered between psychological attributes such as personality traits, attitude, and group performance.
Of the five dimensions of the widely established Big Five personality traits [24], it is primarily extraversion and conscientiousness that are considered relevant to the group formation [13]. For extraversion – a trait implying not only sociability but also leadership behavior as one of its facets – a heterogeneous group composition is deemed to be conducive so as to avoid conflicts between numerous individuals with leadership claim. For conscientiousness, a homogeneous group composition is regarded as beneficial, as larger differences between group members would be detrimental to the formulation of common group goals; Members with greater conscientiousness would probably have higher expectations regarding group performance than would be the case for members of lower conscientiousness. For one individual group, it would also be preferable for conscientiousness within the group to be not only homogeneous, but also high on average, as conscientiousness of individuals is considered a good predictor for higher achievement [26]. Nevertheless, given population that needs to be entirely divided into groups, the average value of one group can only be increased at the expense of the other groups. For reasons of fairness, therefore, group level cannot be optimized.
Beyond the Big Five personality dimensions, various other psychological traits were observed to be relevant to group formation. Bell [2] identifies team orientation as a criterion that should be distributed homogeneously, so that group members may agree more readily on the degree of collaboration. Further discussed factors are motivation and goal orientation [21], as well as general cognitive abilities and previous knowledge [12].
2.2 Computer Supported Learning Group Formation (CSLGF)
Finding optimal solutions with algorithms exceeds practical computability already with course sizes below 40 students. The computational problem belongs to complexity class
Existing algorithmic approaches vary in used methods, the amount of supported criteria types, the supported optimization goals per criterion (e. g. homogeneous, heterogeneous or mixed), and the frequency of group formation (once, repeated). Figure 1 contains an overview illustrating the categorization (extended version based on [29]).
The constraint-based approaches (categorized in Figure 1 on the left) formalize the problem as a Constraint Satisfaction Problem (CSP). Based on rules, which describe valid learning groups, and based on learner characteristics they use semantic technologies – preferably ontologies – to find valid solutions. Ounnas, Davis, and Millard [22] define an extended friend-of-a-friend ontology and support restrictions about group roles, gender and group size which are fed into a common logical problem solver (DLV). Inaba, Supnithi, Ikeda, Mizoguchi, and Toyoda [14] define a model of participants‘ learning goals and apply an agent-based system. Likewise, by using software agents, Abnar, Orooji, and Taghiyareh [1] use the participants’ competencies and rating of (preceding) group work. One issue of agent-based systems is not considering a uniform group quality among all groups, since software agents optimize towards goals of individuals. All semantic approaches have in common the dependency of an ontology supporting a certain set of criteria and learning scenarios. This limits flexibility in changing the set of used criteria or the ability to apply the solution to different learning scenarios. Additionally, these algorithmic approaches do not provide a quality metric (e. g. fitness function) to allow easy comparability of groups and group formation quality of different algorithmic iterations.
The numerical approaches (categorized in Figure 1 on the left) formalize characteristics of learners and context-dependent aspects as n-dimensional criteria vectors. Consequently, this results in a multi-dimensional vector space of all characteristics to be considered in group formation. The two algorithms of Paredes [23] and Christodoulopoulos and Papanikolaou [5] use cluster analysis for group formation, but do not allow optimization for heterogeneity in one criterion and homogeneity in another simultaneously. Graf and Bekele [7] use ant colony optimization to primarily achieve good results in formation of heterogeneous groups. OmadoGenesis [6] allows formation of homogeneous, heterogeneous, and mixed groups, but only supports discrete criteria values and uses for mixed criteria sets (no more than) a genetic algorithm. Similar restrictions concerning the supported diversity of criteria apply to the heuristic, iterative method of Cavanaugh and Ellis [4]. Only two of the approaches found in the literature research provide original algorithms to achieve near optimal solutions by goal-oriented pre-sorting of participants and selection of next group members based on pairwise comparison or comparison with pivot elements [7], [17]. Of these two, GroupAL already proved to deliver better group formation results in comparison to other algorithms, and thus was used as the basis for the Moodle-Plugin and evaluation study presented in the following sections.
In terms of formation frequency (categorized in Figure 1 on the right) the approaches discussed so far primarily aim at near optimal formation results on first (and probably only) algorithmic iteration. A multitude of contributions exists, which focuses on incremental improvements of group formation quality from one group work to the next (e. g. in MOOCs with group work intervals or courses with weekly homework). Ideally, without any prior data about the participants, they build optimized learning groups for the next iteration (solely) based on data about participants behavior, interaction, feedback and the quality of group work (cf. [31], [29]).
2.3 MoodlePeers: Data Collection and Algorithmic Learning Group Formation with GroupAL
In previous work [16], [27] a plugin supporting algorithmic learning group composition for the learning management system Moodle was introduced. The plugin is called MoodlePeers[2] (or mod_groupformation[3] as in the Moodle plugin repository). It is an activity-based plugin which can be used in the course context of Moodle. The plugin was created in an agile development process and was released as an Open Source project licensed under GNU General Public License.[4] In its current version, it supports Moodle with version 2.9 or higher.
By using MoodlePeers, relevant criteria data can be collected with customizable questionnaires. Instructors can choose between three types of learning group scenarios and respectively the content of the questionnaire and the considered criteria vary. Each type, as depicted in Figure 2, may be used for a specific scenario, such as project teams, homework groups or presentation groups.
Besides the scenario, instructors may also configure the maximum number of groups or the maximum group size. In addition, instructors can configure the activity to ask students about their intended grades or points. This way, optimization pays attention to matching students by similar grading aims.
In some scenarios, instructors want to ask students about their prior knowledge in certain areas. During activity configuration, instructors can add a set of questions which require students to estimate their level of prior knowledge on a range between 0 and 100. This enables the plugin usage in various course scenarios and supports group formation based on the students’ knowledge.
If instructors want to assign a fixed number of topics to all students of a course, they may choose the type ‘presentation groups’. Choosing this type, enforces the instructor to enter a set of topics, which need to be sorted from most to least favorite by the students. The resulting number of groups is equal to the number of defined topics.
All questionnaires are internationally recognized in psychological research for being standardized, reliable, and valid measurement instruments. The collected criteria data include the Big Five personality traits, motivational dimensions and criteria regarding team work as well as demographic data.
After initial configuration of the group formation activity, the questionnaire can be answered by students enrolled in the respective course. The plugin supports an explicit opt-in functionality. Students need to agree to the terms of data collection before submitting answers to rather personal questions, i. e. questions regarding personality traits or personal goals. The terms of data collection also clarify that no instructor has access to the answers of the students and that answers will be deleted after a pre-configured number of days (depending on the plugin configuration during installation).
The plugin supports various answer types, e. g. Likert scale, dropdown single choice. Since the questionnaires are customizable, missing questions can be added manually. By default, the plugin is available in English and German. Figure 3 shows a questionnaire page containing questions with a Likert scale answer format.
When a student finishes the questionnaire and submits the answers, textual and visual feedback about own results for the Big 5 personality traits and motivational dimensions is given. It can be compared to the average values of all students in the course, and later – after group formation – with the students of one’s own learning group. Based on cut-off values, a text-based feedback is given for each criterion. Figure 4 shows the visualization of the category ‘Motivation’.
After all answers are received or the instructor decides to close the activity (which means that no more students can answer the questionnaire), the group formation can be started. By computing criteria vectors based on a students’ answers and taking these vectors as an input, a group formation algorithm can compute groups. The plugin implements the GroupAL algorithm which was introduced by Konert [15]. It is an algorithm based on a non-linear optimization problem and supports homogeneous and heterogeneous criteria, as well as variable number of groups or group sizes. MoodlePeers uses a PHP-based implementation of the algorithm and it runs asynchronously as a cron job on the server.
Depending on the selected learning group scenario and the respective questionnaire, for the relevant criteria vectors are created. Each vectors represent one student and the vector values represent her given answers. GroupAL can be configured to use those criteria homogeneously and/or heterogeneously to generate the most suitable groups. In more detail, the criteria can be weighted independently and therefore, the algorithm can be configured very precisely. Two different matchers can be used for building groups. It can be used a Participant-Centric Matcher (PCM) where at first each group gets assigned a random participant and the remaining participants are assigned iteratively to the best suitable group which is not full yet. Alternatively, by using the Group-Centric Matcher (GCM), each group is successively filled with the most suitable participants.
To assure a high-quality result the algorithm optimizes by two performance indices. The GroupPerformanceIndex (GPI) indicates the quality of a single group, while the CohortPerformanceIndex (CPI) represents the overall quality of all groups or the whole cohort. By optimizing using both indices, the overall quality of both, single groups and the resulting cohort, gets optimized. More details about the performance indices and their usage in GroupAL can be found in Konert [15].
For students who did submit none or incomplete answers to the questionnaire the instructor can choose between grouping them at random or not grouping them at all. This way, inactive users in course setting can be excluded from the learning group formation. So far, the algorithm cannot deal with incomplete data in the questionnaire, which is why instructors can configure the activity such that the students are forced to answer all questions without skipping some.
MoodlePeers uses the generated groups first as a suggestion to the instructor. Figure 5 shows the instructors options to edit, delete or adopt the generated groups. The instructor can then edit the suggested groups if necessary or directly adapt them to the course context and automatically create groups in Moodle. Students then see their assignment to a group and can contact all other group members via the internal messaging system or e-mail. Additionally, students can now also review the average values of their group members in the graphical evaluation.
Since the plugin is an Open Source project anybody can edit the questionnaire or the algorithm configuration. For the later presented studies, the authors used a customized questionnaire and a specific algorithmic setup to realize the research design. The default configuration of questionnaires, criteria and weights in the available versions was proven to have positive impact on satisfaction, motivation and results.
2.4 Pilot Study
Due to good psychometric properties and easily accessible measuring instruments, there are more studies on the influence of extraversion and conscientiousness on the group formation than on other influencing factors. As findings are derived from purely correlative studies – typically groups were studied that were formed either by chance or by choice of the students themselves – this should be considered as a limitation of the empirical results. Study designs of higher quality, especially randomized experimental studies, which allow for conclusions on causal effects, are to date largely missing. One exception to this is a pilot study carried out by the authors of this paper. It demonstrated that a combination of personality traits enhances satisfaction and group performance significantly compared to sheer randomized grouping [27]. However, it was subsequently not possible to assess to which relative amount the individual criteria supported the positive overall effect and whether interactions between criteria occurred.
3 Research Questions and Focus of the Study
The present study systematically examines the impact of extraversion and conscientiousness on quality of collaborative learning. In order to analyze the impact of each of the two criteria separately as well as the interaction between them, we apply a 3×3 experimental design in which the distribution of each criterion is manipulated to be either homogeneous, heterogeneous, or not manipulated (i. e., distributed at random). Extraversion is expected to be beneficial when distributed heterogeneously; conscientiousness is expected to be beneficial when distributed homogeneously. When applying both criteria at once, the algorithm faces greater restrictions to optimize group formation compared to conditions where only one criterion is applied – we therefore expect a negative interaction effect. In order to increase robustness of effects, we include conditions in which criteria are theoretically expected to have a maleficent effect (i. e. homogeneous extraversion and heterogeneous conscientiousness). This is ethically unproblematic as our study takes place in a voluntary preparation course without consequences for further academic grading. We also include conditions that ignore one criterion (i. e. distribution at random) and only manipulate the other; there is even one condition in which both criteria are distributed randomly.
Due to the plentitude of possible analyses provided by this complex experimental design, in the following we only focus on a subset of conditions, namely the four conditions in which both criteria are used simultaneously – leaving out those conditions in which only one is manipulated.
Conscientiousness | ||||
heterogeneous | random | homogeneous | ||
|
||||
Extraversion | heterogeneous | 1 | 2 | 3 |
random | 4 | 5 | 6 | |
homogeneous | 7 | 8 | 9 |
3.1 Hypotheses
Based on the current state of empirical findings, we investigate the following hypotheses:
H1) Groups with a heterogeneous distribution of extraversion report greater satisfaction with group composition and group collaboration, show higher time investment for group work, and achieve better results than groups with a homogeneous distribution of extraversion.
H2) Groups with a homogeneous distribution of conscientiousness report greater satisfaction with group composition and group collaboration, show higher time investment for group work, and achieve better results than groups with a heterogeneous distribution of conscientiousness.
3.2 Study Design and Method
A two-factorial study design was used for the analysis of the research questions. Both the factor of extraversion and conscientiousness were subdivided into three categories with the homogeneous, random and heterogeneous distributions. The nine experimental conditions resulting thereof are shown in Table 1. The advantage of this completely crossed experiment design is that the respective influence of extraversion (main effect 1) and conscientiousness (main effect 2) can be examined independently from each other. In addition, a possible interaction effect of the two factors can be tested. In the following, only conditions 1, 3, 7 and 9 are examined and compared with each other.
Conscientiousness | ||||
heterogeneous | random | homogeneous | ||
|
||||
Extraversion | heterogeneous |
|
|
|
|
|
|
||
random |
|
|
|
|
|
|
|
||
homogeneous |
|
|
|
|
|
|
|
Condition 1 | Condition 3 | Condition 7 | Condition 9 | |
|
|
|
|
|
Satisfaction | 4,11 (1,79) | 2,46 (1,13) | 3,00 (1,75) | 3,40 (1,51) |
Productivity | 4,00 (1,89) | 1,62 (0,77) | 2,31 (1,99) | 2,30 (1,57) |
Time investment | 3,00 (3,43) | 0,53 (0,92) | 1,00 (1,36) | 0,82 (1,33) |
Completion of assignments | 0,92 (1,10) | 0,45 (0,79) | 0,39 (0,83) | 0,35 (0,81) |
The assignment to the individual conditions was random. Participants were unable to recognize which condition they were assigned to and did not know the subject of the study, and so, selection effects or expectation compliant behavior were excluded. With these characteristics, the experimental study design corresponds to quality level 1+ according to Harbour and Miller [8]. In order to minimize the overlay with effects of personal sympathy and possibly existing friendships, a mathematics preparatory course for first-semester university students of different study programs was chosen as the setting. Participants exclusively participated online via Moodle for four weeks on the voluntary and ungraded course. As attendance was voluntary and ungraded, the assignment to the experimental conditions was considered possible from an ethical perspective. Under different circumstances it would be essential to notify participants about the nature of the study and permit a fair choice of groups, thus weakening the quality of the experiment. Due to the thus far weak empirical foundation work, a division into partly unfavorable group constellations seems reasonable. Only when evidence of performance enhancing or performance impairing group formations has been produced in high-quality studies, subjects may no longer consciously be put in disadvantageous study conditions. Nevertheless, in the scenario presented, all participants had the opportunity to decline their consent to participate in the study (opt-in with textual data protection statement). They were then divided into random groups that were not part of the control groups, aiming for a group size of four members per group; in exceptional cases groups of three members were also allowed. This decision was made after the weighing of two aspects: on the one hand, a group should be able to compensate for the loss of a member, retaining the ability to keep on working; and on the other hand, the total number of groups formed should be maximized. As part of the prep course, the participants received a total of three group assignments, each of which was to be completed over the course of a week. The tasks were developed by mathematical didactics as open modeling tasks, so that problem solving competence was sharpened and collaboration promoted.
In addition to the questionnaire on the Big Five [24], the variables self-regulation competence [3], motivation [25] and self-efficacy [28] and demographic data were collected as control variables via the MoodlePeers plugin. After completion of the preparatory course, the participants answered an evaluation sheet, which included questions on satisfaction with the group composition and self-assessed productivity level of the group (each Likert-scaled from 1 to 6) and on how much time was spent on group work each week. As an objective measure for assessing the quality of the group work, the number of successfully completed group assignments was evaluated.
The study was conducted with the 660 participants of the prep course in September and October 2016, of which 461 students gave their consent for participate and provided complete data for the group formation. The division into the experimental conditions is included in Table 2.
4 Results
As is common with voluntary courses [27], the prep course showed a considerable dropout and not all of the still remaining participants filled in the evaluation survey for the undertaken study. Of the
A 2×2 ANOVA yielded the following results:
There was a significant interaction effect for the satisfaction with the group composition (
For the self-assessed productivity of the group we found a significant main effect for conscientiousness (
For the self-reported time investment, a significant main effect for conscientiousness (
For the successfully completed group assignments a significant main effect for extraversion (
5 Discussion and Conclusion
The present study investigated whether extraversion and conscientiousness should be distributed homogeneously or heterogeneously within learning groups. A randomized experimental design was applied in a voluntary online mathematics preparation course in which participants were asked to work collaboratively on three weekly assignments. Quality of collaborative work was assessed by self-reported satisfaction, productivity, and time investment as well as quantity and quality of results (i. e., group assignments handed in). In the present study, we focus on a subset of conditions of the experiment, namely those with both criteria manipulated simultaneously.
The findings of our study partly confirm our hypotheses, but to a larger extend results are in direct opposition to our expectations. Concerning hypothesis H1, we found groups with extraversion distributed heterogeneously to be more successful in the weekly assignments. However, analyses of the other three dependent variables revealed no further support for our hypothesis, i. e. we found no differences between conditions for satisfaction, self-rated productivity, or time-investment. When interpreting those findings, one should bear in mind that these measures are only self-report data, additionally suffering from a rather high dropout in our sample. Further support for hypothesis H1 can be derived from significant interaction effects: Heterogeneously distributed extraversion showed beneficial effects when combined with heterogeneous distribution of conscientiousness.
Taken together, we interpret our findings to yield support for the hypothesis that extraversion should be distributed heterogeneously. However, the effects are small and could be limited to the specific setting of our study.
Contrary to our expectations, hypothesis H2 has to be rejected. We find no evidence that conscientiousness should be distributed homogeneously within learning groups. Quite reversely, it seems that heterogeneity of conscientiousness has beneficial effects on group work. We found significant main effects for three out of four dependent variables, namely self-reported productivity and time-investment as well as objectively measured quantity of group assignments. Taken together with the interaction effects mentioned above, we conclude that this beneficial effect of heterogeneous conscientiousness can be even increased through heterogeneously distributing extraversion simultaneously. However, as these findings did not correspond to our prior expectations, future research must yield replications before this question can be answered conclusively.
There are several possible reasons why our results are not in line with prior research, including limitations of our study. Although a sample size of more than 200 participants might seem large at first sight, there were only 54 different learning groups when analyzing the four conditions as presented in this paper. This sample size could be regarded as the lower boundary for granting external validity, even more so in case of the questionnaire data (self-reported satisfaction, productivity, and time investment). Objective performance, on the other hand, is not suffering from missing data, as group assignments were either handed in or not – making this measure the most reliable in our study.
Another reason for diverging results could be the peculiarities of our sample. We investigated students that were enrolled in mathematics, computer science, and engineering majors, leaving us with a predominantly male sample. There was little extrinsic motivation to engage in the group assignments as participation was voluntary and did not provide credit points for the respective majors. Learning groups did not collaborate face-to-face but in virtual teams through chats and discussion boards and furthermore collaboration was not very intense, with one task per week that involved at most some two or three hours of communication. In such a scenario, homogeneity or heterogeneity of group members’ personality traits might be less important because contact was so little intensive that group members might not even have recognizes their peers’ personality, let alone differences or similarities compared to themselves. Instead, it might have been more important for productive collaborative work for each group to include at least one single person that initiates and fosters group work. Such a person should probably be both conscientious (so that he or she takes the assignments seriously) and extraverted (so that he or she starts communication with the group). This hypothesis is well in line with our findings: heterogeneous groups have better chances to include persons with high values and the best condition in our experiment seemed to be the combined heterogeneity in both conscientiousness and extraversion.
One last reason why our study provided different findings than prior research might be limitations of the latter. Effects of group composition on performance so far have only been investigated with correlational study designs where the sample was either randomly assigned to learning groups or through self-sorting procedures. In both cases, a lack of experimental manipulation renders it impossible to draw causal inferences. This is the major strength of our approach and we strongly recommend researchers to adapt experimental designs for future studies.
Compared to the findings from our pilot study [16] that used a combination of multiple criteria, we conclude that the positive overall effect could partly be explained through the heterogeneous distribution of extraversion that was applied then. However, a second criterion used in our pilot study was homogeneously distributed conscientiousness which – in the light of our new findings – should not have contributed to successful group work. Future studies should therefore investigate the relative impact of the remaining criteria from our pilot study, namely heterogeneous distribution of prior knowledge, neuroticism, and openness as well as homogeneous distribution of social agreeableness, motivation, and team orientation. Further, we need to develop a better understanding of task or situation specific factors that influence which criteria are most important in which context.
About the authors
Dr Henrik Bellhäuser studied Psychology at the universities of Saarbrücken and Mainz. In his doctoral thesis at Technische Universität Darmstadt, he developed and evaluated a web-based training which aims at fostering self-regulated learning. Currently, he works as a researcher for educational psychology at Johannes Gutenberg-University Mainz and Goethe-University Frankfurt.
Johannes Konert is professor for Web Engineering at Beuth University of Applied Sciences in Berlin. With a scholarship of German Research Foundation (DFG) for the research training group “Feedback-based Quality Management in E-learning” he developed solutions to add peer education to single-player educational games. Supplemented with his insights gained from foundation of a social media company, his research focuses on optimization of digital learning experiences.
M. Sc. Adrienne Müller studied Psychology at the university of Mainz. Since October 2017 she works as a researcher for educational psychology at Johannes Gutenberg-University Mainz and started her dissertation focusing on group composition regarding personality traits within the project of Dr Henrik Bellhäuser.
René Röpke studied Computer Science and IT Security at Technische Universität Darmstadt and in his Master’s thesis he developed an identity management solution to support learning analytics in open learning environments. He is currently pursuing his doctorate at RWTH Aachen University in the Learning Technologies Research Group. His interests are among collaborative learning, multi-touch learning applications and computer science education.
Literature
[1] Abnar, S., Orooji, F., & Taghiyareh, F. (2012). An evolutionary algorithm for forming mixed groups of learners in web based collaborative learning environments. 2012 IEEE International Conference on Technology Enhanced Education (ICTEE), (pp. 1–6). 10.1109/ICTEE.2012.620861.Search in Google Scholar
[2] Bell, S. T. (2007). Deep-level composition variables as predictors of team performance: a meta-analysis. The Journal of Applied Psychology, 92(3), 595–615. 10.1037/0021-9010.92.3.595.Search in Google Scholar PubMed
[3] Bellhäuser, H., Lösch, T., Winter, C., & Schmitz, B. (2016). Applying a web-based training to foster self-regulated learning – Effects of an intervention for large numbers of participants. Internet and Higher Education, 31, 87–100. 10.1016/j.iheduc.2016.07.002.Search in Google Scholar
[4] Cavanaugh, R., & Ellis, M. (2004). Automating the Process of Assigning Students to Cooperative-Learning Teams. Proceedings of the 2004 American Society for Engineering Education Annual Conference & Exposition. Retrieved from http://personal.stevens.edu/~mardis/papers/Cav_Ell_Lay_Ard_asee004.pdf.Search in Google Scholar
[5] Christodoulopoulos, C. E., & Papanikolaou, K. A. (2007). A Group Formation Tool in an E-Learning Context. 19th IEEE International Conference on Tools with Artificial Intelligence (ICTAI 2007), (pp. 117–123). 10.1109/ICTAI.2007.155.Search in Google Scholar
[6] Gogoulou, A., Gouli, E., Boas, G., Liakou, E., & Grigoriadou, M. (2007). Forming Homogeneous, Heterogeneous and Mixed Groups of Learners. In P. Brusilovsky, M. Grigoriadou, & K. Papanikolaou (Eds), Proceedings of Workshop on Personalisation in E-Learning Environments at Individual and Group Level, 11th International Conference on User Modeling (pp. 33–40).Search in Google Scholar
[7] Graf, S., & Bekele, R. (2006). Forming heterogeneous groups for intelligent collaborative learning systems with ant colony optimization. Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 4053 LNCS, (pp. 217–226). 10.1007/11774303_22.Search in Google Scholar
[8] Harbour, R., & Miller, J. (2001). A new system for grading recommendations in evidence based guidelines. British Medical Journal, 323 (August), 334–336. 10.1136/bmj.323.7308.334.Search in Google Scholar PubMed PubMed Central
[9] Harrison, D. A., Price, K. H., Gavin, J. H., & Florey, A. T. (2002). Time, Teams, and Task Performance: Changing Effects of Surface- and Deep-Level Diversity on Group Functioning. Academy of Management Journal, 45(5), 1029–1045. 10.2307/3069328.Search in Google Scholar
[10] Henry, T. R. (2013). Creating effective student groups. Proceeding of the 44th ACM Technical Symposium on Computer Science Education – SIGCSE 13, 645. 10.1145/2445196.2445387.Search in Google Scholar
[11] Hoppe, U. (1995). The Use of Multiple Student Modeling to Parameterize Group Learning. In J. Greer (Ed.), 7th World Conference on Artificial Intelligence in Education (pp. 234–249). AACE: Charlottesville, VA. Retrieved from http://www.collide.info/Members/admin/publications/AIED95_Hoppe.pdf.Search in Google Scholar
[12] Horwitz, S. K. (2005). The Compositional Impact of Team Diversity on Performance: Theoretical Considerations. Human Resource Development Review, 4(2), 219–245. 10.1177/1534484305275847.Search in Google Scholar
[13] Humphrey, S. E., Hollenbeck, J. R., Meyer, C. J., & Ilgen, D. R. (2007). Trait configurations in self-managed teams: a conceptual examination of the use of seeding for maximizing and minimizing trait variance in teams. The Journal of Applied Psychology, 92(3), 885–892. 10.1037/0021-9010.92.3.885.Search in Google Scholar PubMed
[14] Inaba, A., Supnithi, T., Ikeda, M., Mizoguchi, R., & Toyoda, J. (2000). How Can We Form Effective Collaborative Learning Groups? In 5th International Conference on ITS (pp. 282–291). Springer: Montreal, Canada. Retrieved from http://www.springerlink.com/index/00XNX7Y4T7UYA1VQ.pdf.Search in Google Scholar
[15] Konert, J. (2014). Interactive Multimedia Learning: Using Social Media for Peer Education in Single-Player Educational Games. Springer: Heidelberg, Germany. Retrieved from http://www.springer.com/engineering/signals/book/978-3-319-10255-9.Search in Google Scholar
[16] Konert, J., Bellhäuser, H., Röpke, R., Gallwas, E., & Zucik, A. (2016). MoodlePeers: Factors relevant in learning group formation for improved learning outcomes, satisfaction and commitment in E-learning scenarios using GroupAL. In K. Verbert, M. Sharples, & T. Klobucar (Eds), Adaptive and Adaptable Learning: Proc. of the 11th European Conf. on Techn. Enhanced Learning (EC-TEL 2016) (pp. 390–396). Springer LNCS: Lyon, France. 10.1007/978-3-319-45153-4_32.Search in Google Scholar
[17] Konert, J., Burlak, D., & Steinmetz, R. (2014). The Group Formation Problem: An Algorithmic Approach to Learning Group Formation. In C. Rensing, S. de Freitas, T. Ley, & P. J. Muñoz-Merino (Eds), Proceedings of the 9th European Conference on Technology Enhanced Learning (EC-TEL) (pp. 221–234). Springer Berlin: Graz, Austria. 10.1007/978-3-319-11200-8_17.Search in Google Scholar
[18] Kyndt, E., Raes, E., Lismont, B., Timmers, F., Cascallar, E., & Dochy, F. (2013). A meta-analysis of the effects of face-to-face cooperative learning. Do recent studies falsify or verify earlier findings? Educational Research Review, 10, 133–149. 10.1016/j.edurev.2013.02.002.Search in Google Scholar
[19] McCrae, R. R., & Costa, P. T. (1994). The Stability of Personality: Observations and Evaluations. Current Directions in Psychological Science, 3(6), 173–175. 10.1111/1467-8721.ep10770693.Search in Google Scholar
[20] Mitchell, S. N., Reilly, R., Bramwell, F. G., & Lilly, F. (2012). Friendship and Choosing Groupmates: Preferences for Teacher-selected vs. Student-selected Groupings in High School Science Classes. Journal of Instructional Psychology, 31(1), 1–6. Retrieved from http://web.centre.edu/plummer/readings/228readings/mitchell.pdf.Search in Google Scholar
[21] Nederveen Pieterse, A., van Knippenberg, D., & van Ginkel, W. P. (2011). Diversity in goal orientation, team reflexivity, and team performance. Organizational Behavior and Human Decision Processes, 114(2), 153–164. 10.1016/j.obhdp.2010.11.003.Search in Google Scholar
[22] Ounnas, A., Davis, H., & Millard, D. (2008). A Framework for Semantic Group Formation. Eighth IEEE International Conference on Advanced Learning Technologies, 34–38. 10.1109/ICALT.2008.226.Search in Google Scholar
[23] Paredes, P. (2010). A Method for Supporting Heterogeneous-Group Formation through Heuristics and Visualization, 16(19), 2882–2901.Search in Google Scholar
[24] Rammstedt, B., & John, O. P. (2005). Kurzversion des Big Five Inventory (BFI-K): Entwicklung und Validierung eines ökonomischen Inventars zur Erfassung der fünf Faktoren der Persönlichkeit. Diagnostica, 51(4), 195–206. 10.1026/0012-1924.51.4.195.Search in Google Scholar
[25] Rheinberg, F., Vollmeyer, R., & Burns, B. D. (2001). FAM: Ein Fragebogen zur Erfassung aktueller Motivation QCM: A questionnaire to assess current motivation in learning situations. Diagnostica, (47), 57–66.Search in Google Scholar
[26] Richardson, M., Abraham, C., & Bond, R. (2012). Psychological correlates of university students’ academic performance: a systematic review and meta-analysis. Psychological Bulletin, 138(2), 353–387. 10.1037/a0026838.Search in Google Scholar PubMed
[27] Röpke, R., Gallwas, E., Konert, J., & Bellhäuser, H. (2016). MoodlePeers: Automatisierte Lerngruppenbildung auf Grundlage psychologischer Merkmalsausprägungen in E-Learning-Systemen. In U. Lucke, A. Schwill, & R. Zender (Eds), Proc. der 14. E-Learning Fachtagung Informatik der g.i. (DeLFI 2016) (pp. 233–244). Köllen Druck+Verlag GmbH.: Bonn.Search in Google Scholar
[28] Schwarzer, R., & Jerusalem, M. (1999). Skalen zur Erfassung von Lehrer- und Schülermerkmalen. Dokumentation der psychometrischen Verfahren im Rahmen der Wissenschaftlichen Begleitung des Modellversuchs Selbstwirksame Schulen. Verfahren im Rahmen der (Vol. 38).Search in Google Scholar
[29] Srba, I., & Bielikova, M. (2014b). Dynamic Group Formation as an Approach to Collaborative Learning Support. IEEE Transactions on Learning Technologies, 8(99), 173–186. 10.1109/TLT.2014.2373374.Search in Google Scholar
[30] Wessner, M., & Pfister, H.-R. (2001). Group formation in computer-supported collaborative learning. In Proceedings of the 2001 International ACM SIGGROUP Conference on Supporting Group Work – GROUP’01 (pp. 24–31). 10.1145/500286.500293.Search in Google Scholar
[31] Zheng, Z. (2013). A Dynamic Group Composition Method to Refine Collaborative Learning Group Formation. In Proceedings of the 6th International Conference on Educational Data Mining (EDM) (pp. 360–361).Search in Google Scholar
© 2018 Walter de Gruyter GmbH, Berlin/Boston