Adapting Online Group Formation to Learners’ Conscientiousness, Agreeableness and Ability Chinasa Odo∗ Judith Mastho� Nigel Beacham r01cro17@abdn.ac.uk Utrecht University and University of Aberdeen University of Aberdeen University of Aberdeen Aberdeen, UK Aberdeen, UK Utrecht, the Netherlands ABSTRACT becomes a factor in facilitating collaborative learning. These op- This paper focuses on the impact of conscientiousness, agreeable- portunities are accompanied by approaches that emphasize the ness, and ability for the formation of heterogeneous learning groups learner as the main agent of learning. Learners in this situation in supporting lifelong learning. It presents a study in which partici- come together to make learning socially interactive rather than a pants assigned learners to groups to investigate whether these, and transmission of pre-packaged lectures [16]. When learners engage more importantly, how they use learner personality and ability in in online collaborative learning, it may help to induce positive a�ect group formation and inspire future algorithms. by providing an opportunity for active participation in achieving the learning objectives within the group [4]. Collaborative learning CCS CONCEPTS is a situation where two or more learners come together to facilitate learning [11]. The aim is to provide learning activities that give • Human-centered computing → Collaborative and social learners opportunities to interact, share and process information. A computing. collaboration paradigm promotes problem-solving, critical thinking and facilitates the development of interaction between learners [49] KEYWORDS and promotes an overall participation of all learners [5]. Group formation, Collaborative learning, User study, Adaptation The collaborative environment enables teachers to be facilita- tors who assist in generating and sharing learning content, and not to control the delivery and pace of learning [42]. Teachers en- 1 INTRODUCTION sure that the core concepts and practices of the subject domain are Learning is seen as a continuous process which starts from birth fully integrated, and are also responsible for creating the environ- and terminates at death. According to the Commission of the Eu- ment through which e�ective collaboration can be possible [40]. ropean Communities [8], lifelong learning is de�ned as all forms A learner can engage in discussions in which they construct and of voluntary or self-motivated learning undertaken by adults after share the understanding of content through di�erent methods [21]. their initial education and training. Lifelong learning encompasses This is inspired by the Zone of Proximal Development by Vygot- continuing education and professional development programs for sky [47]. Vygotsky believes that any learning encounters have a self-sustainability, competitiveness and employability. The aim is previous history, and he emphasized the importance of learning to support individuals to remain relevant in the �eld, since it is through interactions with others rather than individual work. Sup- not possible to acquire all the required knowledge during the tra- porting Vygotsky is the cognitive developmental theory of Piaget ditional school years. With the emergence of technology, lifelong [41] which noted that the cognitive development is a progressive learning has no barrier on how we receive and gather information, transformation of mental developments caused by biological ad- collaborate and communicate with others. According to Laal [25], vancement and those acquired within the environment. In a social lifelong learning is diverse, adapted to individuals and available learning system also, Bandura [2] noted that new patterns of be- throughout life unlike traditional learning. Lifelong learning is of- haviour can be acquired through direct experience or by observing ten not teacher lead in contrast to traditional classroom learning, the behaviour of others. The interaction within the online learning but individual learners can collaborate with others to enhance their environment may induce positive changes in learners’ a�ective understanding and skills. state. Traditionally, teachers have been the leading source of knowl- This paper investigates automatic group formation, to improve edge transmissions in the learning environment. As technology the e�ectiveness of collaboration in supporting the learning process becomes more advanced, the opportunity provided by e-learning within and beyond the walls of educational system. In particular it investigates to what extent learners’ conscientiousness, agree- Permission to make digital or hard copies of all or part of this work for personal or ableness and ability should inform group formation and in which classroom Copyright use is granted held without by the fee provided author(s). Usethat copies areunder permitted not made theorCC-BY distributed way. forlicense pro�t orCreativeCommons.org/licenses/by/4.0/ commercial advantage and that copies bear this notice and the full citation on the �rst page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior speci�c permission 2 RELATED WORK and/or a fee. Request permissions from permissions@acm.org. AIED 2019, Chicago, USA, 10.1145/1122445.1122456 2.1 Online Collaboration © 2018 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 978-1-4503-9999-9/18/06. . . $15.00 Online collaboration, just like a conventional group, is formed when https://doi.org/10.1145/nnnnnnn.nnnnnnn two or more people interact and in�uence each other’s discussion AIED 2019, Chicago, USA, 10.1145/1122445.1122456 Odo, et al. for the purpose of learning and understanding learning contents to gender and ability, is an important factor. Also, a study by Davis completely [11, 12]. In an online collaboration, the discussion is [9] suggested a random selection to exploit heterogeneity, to have a central to learning. It operates in an environment that may be mix of males and females, verbal and quiet, the cynical and the op- asynchronous and is independent of place. Through collaborations, timistic learners in a group. Cen et al. [7] found that heterogeneous learning is simpli�ed, because members will strive to motivate groups with a diversity of skills and genders bene�t more from and support one another through discussion and elaboration on collaborative learning than homogeneous groups. Several studies learning activities. The theory of online collaborative learning by have been conducted on personality in group formation [15, 45]. Harasim [17] believed that learners solve problems collaboratively Learners felt that in addition to ability, other learner characteris- through discourse rather than recite what they think is the right tics such as personality needed to be considered. Based on this work, answer. The social environment in�uences the learning process we decided to investigate alternative solutions to group formation [27] and enables learning to be achieved through the process of using computational methods. We conducted a systematic literature observational learning [3]. The social learning theory of Bandura review on group formation for collaborative learning [37] which [2] emphasizes the role of environmental in�uence in learning. It investigated which learner characteristics are used in (automatic) is imperative that learners with the same cognitive characteristics group formation, and what algorithms are used to do the group be matched together to promote learning and foster e�ective team formation. We found that the reviewed papers did not speci�cally performance [1]. consider which learner characteristics are important when forming a group, but tended to focus on a particular characteristic for group 2.2 Group Formation formation. Learning characteristics used included gender, learning style, interests, ability, knowledge, and personality. A variety of Bringing learners together to form a group in terms of ability and methods were used for automatic group formation (see [38] for de- experience has been found to have a positive e�ect on performance tails). The reviewed papers did not base their algorithms on studies [33]. Group roles are largely dependent on learner’s personality and with humans, nor did they evaluate the algorithms with humans. experience [36]. Nazzaro and Strazzabosco [36] observed that some The work in this paper extends the related work by investigating learners are shy, some are impatient while some are con�dent, but in more detail which combinations of learner characteristics to use, what matters in a group is communication among members of the and in which way, based on a study with learners. group. According to Sherif and Sherif [43], groups are constituted to provide individuals with mutual support and the opportunity to solve di�cult problems. Groups are also formed to bring together di�erent characteristics of individuals [11]. The aim is to have a 2.4 Personality and Collaboration good blend of learners who would share ideas to achieve optimal Personality traits are habitual patterns of behavior, thought, and learning outcomes [14]. Moreland et al. [35] regard group compo- emotion that are relatively stable over time, di�er across individ- sition as a cause that can in�uence other aspects of group life, for uals, are consistent over situations, and in�uence behavior [24]. instance, group structure, dynamics and performance. The study by A good combination of personality traits may harness individual Vrioni [46] shows that group learning provides an opportunity for strengths and manage the weaknesses toward a common goal. Good negotiation which creates an environment necessary for learning. and bad personality traits within a team may o�set one another Moreland et al. [35] believe that if a group is e�ectively created, an and build on each other and lead to synergies. Personality traits ideal group can be formed where learners can work together for are an important aspect in group formation because, when person- the optimal satisfaction of a set of learning objectives. A study by ality di�erences are ignored, a team may not perform e�ectively. Odo et al. [39] advocated for learners’ a�ective state being taking The theory in [13] supported team formation and describes how into account when forming collaborative groups. individual personalities interact at the group level. Individuals that di�er in their personality traits exert various in�uences on group 2.3 Heterogeneous Groups behaviour. This is supported by Lykourentzou et al. [29] who de- For groups to be heterogeneous, a distribution is needed of learners �ned balanced groups as consisting of individuals with compatible over the groups which provides diversity, for example in age, gender, personalities. The Myers Briggs Type Indicator is one tool to iden- abilities, skills, cultural background, personality traits, etc, rather tify individual personality type, related to the communication and than the same characteristics being together in one group [48]. interaction within a group [20]. Personality traits in�uence the way According to Houldsworth and Mathews [18], group performance individuals perceive, plan, and execute any activities [6]. Another is in�uenced by the degree of heterogeneity in formation. They type of personality model that exists is the so-called Big-�ve. This found that diverse groups perform more consistently. However, model distinguishes �ve distinct personality dimensions: agreeable- they noted that most groups possess certain elements of process ness, conscientiousness, extroversion, neuroticism and openness to loss as well as aspects of process gain which often tend to balance experience [22]. Understanding personality types when forming a each other out as the group progresses. Considering the hetero- collaborative group may be helpful to appreciate that people are geneous aspect of group composition, a study by Moreland [34] di�erent, with values, special strengths and qualities, and should suggested that age, gender, and cultural background should be be treated with care and respect. Personality traits are important considered as the most important demographic factors for group determinants of human behaviour [22] and may therefore impact formation. Supporting [34] are studies by Jackson et al. [19] and collaboration. McGivney et al. [32] noted that a combination of Lai [26], which maintained that group composition, with respect di�erent personalities impacts group performance and interaction. Adapting Online Group Formation AIED 2019, Chicago, USA, 10.1145/1122445.1122456 3 STUDY: USER-AS-WIZARD RQ2 Is agreeableness considered in group formation, and if so The related work and our own earlier qualitative work [30] (we how? conducted a survey and focus groups) indicated that teachers and RQ3 Is ability considered in group formation, and if so how? students felt personality traits and ability need to be considered For each of these three learner characteristics (CHAR), we are in group formation. However, these studies only provide peoples interested to know: perceptions on what they consider important, and do not show (1) Are high CHAR learners distributed evenly across the groups? what people actually do when forming groups, assuming that the (2) Are low CHAR learners distributed evenly across the groups? learner characteristics are known to them. This study investigates (3) Are individual groups balanced on CHAR, so is the number actual group formation. Many personality traits exist; here we focus of low and high CHAR learners the same? on just two of these, namely conscientiousness and agreeableness, (4) Is CHAR cohesion in individual groups considered? both of which seem relevant to collaboration. The intention is to This results in research questions RQ1.1 RQ1.4, RQ2.1 RQ2.4, and repeat the study with other personality traits in future. RQ3.1 RQ3.4. Regarding even distributions, given there are 3 high CONS and 4 3.1 Design high ABLE learners, an even distribution for high CONS and high This study used the user-as-wizard method [30], in which partici- ABLE means when creating: pants took the part of the adaptive system and had to assign learners • 3 groups of 4: 1 high CONS learner per group; at least 1 high to groups1 . Twenty four participants took part in the study, all were ABLE learner per group students with experience of group learning (10 undergraduates • 4 groups of 3: 1 high ABLE learner per group; no more than and 14 postgraduates; 6 of the postgraduates had also worked as 1 high CONS learner per group teaching assistants and been involved in forming groups of stu- • 2 groups of 6: 2 high ABLE learner per group; no more than dents to work together in a project). They were presented with 12 2 high CONS learners per group learners and their individual learner characteristics and told that The high CONS case is similar for low CONS, high AGR, and low these learners di�ered in personality and ability. They were asked AGR; and the high ABLE case is similar to the low ABLE case. to put these learners into groups, in such a way that the groups Regarding cohesiveness, we believe that a group has better cohe- would work well together2 . sion when the standard deviation of the group’s CHAR is smaller. All learners had common English male names, selected to avoid We calculate the standard deviation by coding high CHAR as 2, any in�uence of gender, ethnicity or religion. Three learner charac- medium 1, and low 0. This for example means that a group of 3 teristics were used: ability (high, low, average), conscientiousness high and 1 low CHAR learners has worse cohesion than a group of (high, low, and average), and agreeableness (high, low, average). 2 high, 1 medium, and 1 low CHAR learners, which has worse co- Validated stories of personality traits of �ctitious learners [10] were hesion than a group of 1 high, 1 low, and 2 medium CHAR learners. used to illustrate the personality traits (four stories depicting high and low levels of conscientiousness and agreeableness). These sto- Table 1: Learner characteristics and personality stories used ries were shown by Smith et al. [10] to reliably convey personality Personality trait Stories types, so we can be con�dent that the participants will interpret High Charles, George, Kenneth. For example: Charles is al- ways prepared. He gets tasks done right away, paying the personality traits correctly. attention to detail. He makes plans and sticks to them Table 1 shows the learner characteristics and personality stories and carries them out. He completes tasks successfully, used and ability levels. We will use the following abbreviations: doing things according to a plan. He is exacting in his work; he �nishes what he starts. ABLE = ability, CONS = conscientiousness, AGR = agreeableness. Conscientiousness Low David, Henry, Larry. For example: David procrastinates Each participant �rst assigned the 12 learners to 3 groups of 4 and wastes his time. He �nds it di�cult to get down to learners, next to 4 groups of 3 learners, and �nally to 2 groups of 6 work. He does just enough work to get by and often doesn‘t see things through, leaving them un�nished. learners. He shirks his duties and messes things up. He doesn‘t Whilst the literature and our previous work indicates that per- put his mind on the task at hand and needs a push to get started. sonality and ability are perceived to matter when forming groups, Medium Anthony, Brian, Edward, Frank, Ian, James we wanted to better understand how these characteristics are used High Anthony, Edward, Ian. For example: Anthony has a when forming groups, in order to be able to produce an algorithm good word for everyone, believing that they have good intentions. He respects others and accepts people as for doing this automatically. So, we were not just interested in they are. He makes people feel at ease. He is concerned whether learner characteristics matter, but particularly in how they about others, and trusts what they say. He sympathizes matter. Hence, we investigated the following overarching research with others‘ feelings and treats everyone equally. He is easy to satisfy. questions: Agreeableness Low Brian, Frank, James. For example: Brian has a sharp RQ1 Is conscientiousness considered in group formation, and if tongue and cuts others to pieces. He suspects hidden motives in people. He holds grudges and gets back at so how? others. He insults and contradicts people, believing he 1 Using this method for this purpose has limitations. These and the rationale for doing is better than them. He makes demands on others and is out for his own personal gain. this anyway will be discussed in the paper conclusions. Medium Charles, David, George, Henry, Kenneth, Larry 2 The instruction to participants was generic on purpose. We did not ask them to make High Anthony, Brian, George, Henry a high performing group, as they may have disregarded the learning outcomes for the Ability Low Charles, David, Ian, James other groups. We did not ask them to make well-balanced groups with approximately Average Edward, Frank, Kenneth, Larry equal conditions AIED 2019, Chicago, USA, 10.1145/1122445.1122456 Odo, et al. 3.2 Results Tables 2 and 3 show the groups created in terms of personality traits and ability respectively. For example, Table 2 shows that when Table 2: Personalities traits combined when forming collab- allocating the learners to 3 groups of 4 learners, 15 participants put orative groups (Conscientiousness and Agreeableness) a high CONS, low CONS, high AGR, and low AGR learner in the Conscientiousness Agreeableness Group �rst group they created. High Low High Low 1 2 3 4 1 1 1 1 15 10 8 RQ1 Is conscientiousness considered in group formation? Partici- 1 1 2 - 7 1 - pants clearly took CONS into account. 3 groups of 4 learners 1 - 1 2 - 7 2 (1) 3 groups of 4. Regarding RQ1.1, only 3 groups (out of 72) 1 2 - 1 - 2 3 1 1 - 2 2 - 3 were created that did not contain a high CONS learner, show- 1 - 2 1 - 1 - ing participants distributed the 3 high CONS learners quite 2 1 - 1 - 3 - evenly over the groups. Regarding RQ1.2, only 10 groups did 1 2 1 - - - 5 not contain a low CONS learner, so also low CONS learners - 1 3 - - - 1 - 1 2 1 - - 2 tended to be distributed, but given the higher number of 1 1 - 1 9 6 - 4 groups without a low CONS learner, it seems participants 1 - 1 1 2 7 6 - felt it was more important that a group contained a high 4 groups of 3 learners - 1 2 - 3 2 - - CONS learner than that the low CONS learners were evenly 1 1 1 - 8 - 3 4 distributed. Regarding RQ1.3, there was no balance of CONS - - 1 2 2 3 2 1 3 - - - - 2 - - in 26 groups, so balance does not seem to be an important - - 2 1 - 2 - 1 consideration for CONS. Regarding RQ1.4, all groups created 1 - 2 - - 1 4 - had good CONS cohesion; there were no groups combining 1 - - 2 - 1 - 2 3 high with 1 low CONS, or 2 high with 2 low CONS. - 1 1 1 - - 4 5 2 1 - - - - 2 - (2) 4 groups of 3. Regarding RQ1.1, only 4 groups (out of 96) were 1 2 - - - - 3 2 created that contained more than one high CONS learner, - 2 - 1 - - - 5 again showing that participants tried to distribute these - 2 1 - - - - 1 evenly. Regarding RQ1.2, there were 11 groups with more 2 2 1 1 7 2 than one low CONS, again showing that high CONS was 2 groups of 6 learners 1 2 2 1 3 - 1 2 1 2 8 1 deemed more important than low CONS when balancing 1 1 2 2 2 7 groups. Regarding RQ1.3, only 27 groups were balanced, so 1 1 1 3 3 - balance does not seem an important consideration for CONS. 2 1 2 1 1 8 Regarding RQ1.4, with a group of 3, the worst cohesiveness 2 1 1 2 - 3 2 2 2 - - 3 is when 2 high CONS are combined with 1 low CONS, or the other way around. This only happened in 7 groups, so cohesiveness was �ne. (3) 2 groups of 6. Regarding RQ1.1 and RQ1.2, only groups were Table 3: Learner ability combined when forming groups created that contained at least one high and one low CONS Ability Group learner. This con�rms that high and low CONS learners High Low Average 1 2 3 4 were distributed evenly over the groups. Regarding RQ1.3 1 2 1 11 1 4 and RQ1.4, half the participants allocated 2 high and 2 low 2 2 - 4 6 - CONS to the same group, seemingly trying to fully balance 2 1 1 9 11 - out the CONS levels across groups, now this now longer had 3 groups of 4 learners 1 1 2 - 4 11 a big impact on CONS cohesion (as the group size meant 2 - 2 - 1 1 there were 2 medium CONS learners in those groups as well). - 1 3 - - 5 So overall, CONS was considered, and in particular high CONS 1 - 3 - - 1 learners are distributed evenly. CONS cohesion is important, and - - 4 - - 1 - 2 2 - 1 1 CONS balance is only considered when it does not impact CONS 1 1 1 9 17 10 11 cohesion. The impact of CONS on group formation is not surprising, 1 2 - 6 1 3 1 because as noted by [28], conscientiousness helps one to ensure 2 1 - 4 - 1 3 and maintain harmonious relationships with others in the group. 4 groups of 3 learners This is because conscientious people are usually well organized, - 2 1 4 - - 3 - 1 2 1 - 3 1 prudent, thorough, neat and achievement oriented [31]. 2 - 1 - 4 - - 1 - 2 - 2 7 5 RQ2 Is agreeableness considered in group formation? AGR is clearly 2 2 2 13 13 less considered than CONS when forming groups. 2 3 1 3 6 (1) 3 groups of 4. Regarding RQ2.1, 13 groups (out of 72) did 2 1 3 6 3 not contain a high AGR learner, showing that participants 2 groups of 6 learners 1 4 1 2 - paid more attention to evenly distributing high CONS across 3 - 3 - 2 Adapting Online Group Formation AIED 2019, Chicago, USA, 10.1145/1122445.1122456 groups than high AGR. Regarding RQ2.2, 16 groups did not e�ort was taken to evenly divide the high CONS learners. contain a low AGR learner. Regarding RQ2.3, only 31 groups Regarding RQ3.2, 18 groups did not contain a low ABLE were balanced on AGR. Regarding RQ2.4, all groups created learner. This result is quite similar to that for CONS. Regard- had good AGR cohesion; there were no groups combining 3 ing RQ3.3, only about half the groups (47) were balanced on high with 1 low AGR, or 2 high with 2 low AGR. ABLE, so this was less of a concern than it seems to have (2) 4 groups of 3. Regarding RQ2.1, despite there not being enough been for the smaller groups. Regarding RQ3.4, there were high AGR learners to allocate even one to each group, in 13 19 groups containing 2 low and 1 high ABLE learner or the groups more than one high AGR leaner was allocated. This other way around, so cohesiveness was not that good. provides evidence that many participants were not trying (3) 2 groups of 6. Regarding RQ3.1, only 4 groups did not contain to evenly distribute high AGR learners across groups. Re- exactly two high ABLE learners, showing that participants garding RQ2.3, they were also not trying to balance out AGR allocated the high ABLE learners evenly over the groups. In within groups, as none of these groups with two high AGR contrast, regarding RQ3.2, 22 groups did not contain exactly learners was allocated two low AGR learners. Regarding two low ABLE learners, showing that evenly distributing RQ2.2, 9 groups contained more than one low AGR leaner, low ABLE learners was deemed less important. Regarding so low AGR learners were slightly more evenly distributed RQ3.3, most people (13 out of 24) balanced ABLE, allocated than high AGR ones. Regarding RQ2.4, there were 10 groups 2 low and 2 high ABLE learners to each group. Regarding combining 2 high AGR with 1 low AGR or the other way RQ3.4, only 2 groups had very bad cohesiveness, containing around, so cohesion is not as good as for CONS, but still �ne. 4 low, 1 high and 1 average ABLE learners. (3) 2 groups of 6. Regarding RQ2.1 and RQ2.2, all groups con- Overall, there is evidence of ABLE being taken into account, but tained at least one high AGR learner, and only 3 groups (out not necessarily as expected. High and low ABLE learners were not of 48) did not contain a low AGR learner. So, in this case, distributed as evenly as could have been possible, and participants there is more evidence of evenly distributing high AGR learn- seem to have cared more about evenly distributing the high CONS ers than low AGR ones. Regarding RQ2.3, only 18 groups learners than the high ABLE learners. However, the groups were contained the same number of high and low AGR learners, so quite similar in average ability, and the most frequently created there is less evidence of balancing than for CONS. Regarding groups were balanced on ability. Cohesiveness was an issue for the RQ2.4, only 3 groups had bad cohesion, combining 3 low smallest (size 3) groups, where balance seems to have been more with 1 high AGR learners. important, but was good for the larger groups. Most groups created Overall, there is some evidence of AGR being considered, but clearly combined low, average and high ABLE learners, which is in line it is considered less than CONS. Balancing the AGR in a group does with a study of Kardanova and Ivanova [23] who suggested that not seem to be a consideration, but there is some evidence that there needs to be a combination of low, average and high learners’ ARG cohesion matters. For AGR, cohesion seems to matter more ability to maintain good performance. than evenly distributing high and low AGR learners, though there is some evidence of the latter as well. Considering AGR in group formation is supported by the result of Lun and Bond [28] who 4 CONCLUSIONS noted that agreeable persons are more socially accommodating and The success or failure of group collaboration depends on how well thus achieve a higher level of relationship harmony with the others individual learners can work together toward a common goals. This in the group. paper has investigated the impact of personality (conscientiousness and agreeableness) and ability on actual behaviour when forming RQ3 Is ability considered in group formation? ABLE is considered groups. The study showed that personality and ability are taken when forming groups, but less so than one may have expected. into account for group formation and, most importantly, provided (1) 3 groups of 4. Regarding RQ3.1, despite there being more high insights on how they should be taken into account, which can be ABLE learners than groups, there were still 6 groups without used in the design of an algorithm that adapts group formation to a high ABLE leaner. Similarly, regarding RQ3.2, there were learner characteristics. Automated group formation is important still 4 groups that did not contain a low ABLE learner and particularly in a setting where there is no human teacher involved or many groups with 2 low ABLE learners. So, participants where one human teacher is dealing with many learners. With the did not tend to evenly divide high and low ABLE learners advance of lifelong learning, there is a move away from traditional across groups. Regarding RQ3.3, most groups tended to be classrooms and from teacher-led learning. Lifelong learners are as balanced on ability as possible (given there are 4 high and motivated to keep learning and keep collaborating with others 4 low ABLE learners and only 3 groups, many groups had to in order to be current in their professional lives. This continuous have 2 of one type and 1 of the other). Regarding RQ3.4, there collaborative learning process will be easier when there is e�ective were no groups containing 3 high or 3 low ABLE learners, automatic group formation. and cohesiveness was generally �ne. One limitation of this work is that participants were asked to (2) 4 groups of 3. Regarding RQ3.1, given there were 4 high ABLE form �ctional groups, so did not get feedback on how well these learners, one could have been allocated to each group. In 16 groups ended up performing. The work in this paper provides groups (out of 96) this did not happen. Comparing this to initial insights for the algorithm, but further studies are needed the results for CONS, in the case when there were as many to investigate the impact of adaptive group formation on learner high CONS learners as groups (3 groups of 4), clearly more motivation and achievement. AIED 2019, Chicago, USA, 10.1145/1122445.1122456 Odo, et al. A second limitation is that the use of the User-as-Wizard ap- considering other characteristics as secondary or not at all. How- proach presumes that participants are good at the task the system ever, participants did not complain about the task being too di�cult, is supposed to perform, so that the behaviour of participants can and the characteristics participants focused on, and the commonali- be used as a basis for an algorithm. Our participants were students, ties in their approaches, still provide valuable insights. The concern and one could query whether they have enough experience to be about task complexity is the reason why we only considered two able to make good groups, and whether it would have been better to personality traits in this study. use teachers (though some of our sample were in fact also teaching Future work will also include studies on other personality traits, assistants). Our earlier qualitative studies had shown that students more detailed analysis on the interaction between characteristics, and teachers had very similar views on group formation. There and studies evaluating the impact of an adaptive group formation is also not much evidence that teachers are better at this task (in algorithm on the motivation and performance of learner groups fact in our earlier focus groups, students complained that teachers and individual learners. often got groupings wrong). The learners used in this study had A system that performs automated group formation will also more recent experience of what it is like to work in groups than require the relevant learner characteristics, such as learner per- teachers would have3 . However, this does not mean the learners sonality. This paper did not discuss how such characteristics can are necessarily good at this task. We did make the task easier for be obtained. For example, there are many ways to detect learner the participants than it normally would be for teachers, in that we personality; see [44] for a review and for a very easy method to provided detailed information on each learner in terms of their obtain learner personality using personality scales. ability and personality, whilst teachers often may lack in particular the latter. ACKNOWLEDGEMENT Whilst the User-as-Wizard method has this limitation, it was hard to conceive of a better way of gaining the insights we needed. The �rst author’s PhD is supported by TETFund, Federal Republic An alternative method would have been to assign real learners to of Nigeria. groups, and measure group collaboration and learning outcomes. However, this would be extremely di�cult to do in a controlled way. REFERENCES Learners vary on many di�erent characteristics (even in personality, [1] W Allinson and John Hayes. 2012. The Cognitive Style Index. In Technical Manual the most popular Five Factor Model distinguishes �ve traits), so and User Guide. Vol. 33. Pearson Education Ltd or its a�liate, United Kingdom, Chapter 8, 119–135. trying to investigate the in�uence of individual traits as well as [2] Albert Bandura. 1977. Self-e�cacy: Toward a unifying theory of behavioral other characteristics (e.g. ability, gender, ethnicity) with real learn- change. Psychological Review 84, 2 (1977), 191–215. [3] Albert Bandura, Gian Vittorio Caprara, Claudio Barbaranelli, Maria Gerbino, and ers is di�cult. One possibility would be to �nd 12 learners who Concetta Pastorelli. 2003. Role of A�ective Self-Regulatory E�cacy in Diverse varied in exactly the way required for the study, whilst being equal Spheres of Psychosocial Functioning. Child Development 74, 3 (2003), 769–780. on all other characteristics (such as gender, ethnicity, and other [4] Nigel Beacham and Bob Duncan. 2017. Development of a Secure Cloud Based Learning Environment for Inclusive Practice in Mainstream Education. Cloud personality traits), which would be hard. We would also require Computing 2017: The Eighth International Conference on Cloud Computing, GRIDs, multiple sets of such learners, to investigate the impact of di�erent and Virtualization 2, 1 (2017), 1–4. groupings. Another possibility would be use a very large set of [5] Nigel Beacham and Martyn Rouse. 2012. Student teachers’ attitudes and beliefs about inclusion and inclusive practice. Journal of Research in Special Educational learners, allocate them to groups randomly, and use data analytics Needs 12, 1 (jan 2012), 3–11. to determine the characteristics of the groups that performed best. [6] Luiz Fernando Capretz and Faheem Ahmed. 2010. Why Do We Need Personality Diversity in Software Engineering? SIGSOFT Softw. Eng. Notes 35, 2 (March 2010), However, the quantity of learners required would be enormous, and 1–11. there could also be an in�uence of the learning task. On balance, [7] Ling Cen, Dymitr Ruta, Leigh Powell, Benjamin Hirsch, and Jason Ng. 2016. Quan- using the User-as-Wizard seems a good approach to investigate titative approach to collaborative learning: performance prediction, individual assessment, and group composition. International Journal of Computer-Supported which learner characteristics learners feel matter, and in which Collaborative Learning 11, 2 (jun 2016), 187–225. way. This narrows down the learner characteristics we ought to [8] Commission of the European Communities. 2009. Adult learning: It is never too consider in future studies with real learners, making it more feasible late to learn. [9] Barbara Gross Davis. 1993. Tools for Teaching (2nd ed.). jossey - Bass Inc., to conduct such studies. Based on the results of this study improved Publishers, 350 Sansome Street, San Francisco, California. 1–449 pages. hypotheses can be formulated, and tested with more participants [10] Matt Dennis. 2006. Adapting Feedback to Personality to Increase Motivation. Technical Report. 7–9 pages. including teachers. [11] Pierre Dillenbourg. 1999. What do you mean by ’collaborative learning’? Americas Another possible limitation is the potential task complexity for Conference on Information Systems 12, 1 (1999), 1–19. study participants. The algorithms previously used for group for- [12] Gloria J. Galanes and Katherine (Katherine L.) Adams. 2013. E�ective group discussion : theory and practice. McGraw-Hill. 443 pages. mation are commonly based on �nding constraint satisfaction so- [13] Jerry W. Gilley, M. Lane Morris, Alina M. Waite, Tabitha Coates, and Abigail lutions [37], which can be considered as a complex task even for Veliquette. 2010. Integrated Theoretical Model for Building E�ective Teams. computers. When letting this task be performed by humans, it Advances in Developing Human Resources 12, 1 (2010), 7–28. [14] Hackman and R Wageman. 2005. When and how team leaders matter. Research is possible that they found it too hard to focus on multiple char- in organizational behavior 26 (2005). acteristics (constraints) at once, which may result in them only [15] Terry Halfhill, Eric Sundstrom, Jessica Lahner, Wilma Calderone, and Tjai M Nielsen. 2005. Group Personality Composition and Group E�ectiveness An Integrative Review of Empirical Research. [16] Edmund J. Hansen and James A. Stephens. 2000. The Ethics of Learner Centered 3 We did not measure the extent of this experience, which would be good to do in Education: Dynamics that Impede the Process. Technical Report. Emporia State future studies of this kind. However, collaborative group work is used a lot in UK University of Kansas, Kansas. 1–8 pages. computing science programmes, including the classes the students were enrolled in, [17] Linda Harasim. 2011. Learning Theory and Online Technologies. In Learning so the students would have had plenty of experience. Theory and Online Technologies. Routledge Taylor & Francis. Adapting Online Group Formation AIED 2019, Chicago, USA, 10.1145/1122445.1122456 [18] Chris Houldsworth and Brian P. Mathews. 2000. Group composition, performance [34] Richard L. Moreland. 1985. Social categorization and the assimilation of "new" and educational attainment. Education + Training 42, 1 (feb 2000), 40–53. group members. Journal of Personality and Social Psychology 48, 5 (1985), 1173– [19] Susan E Jackson, Joan F Brett, Valerie I Sessa, Dawn M Cooper, and Et Al. 1991. 1190. Some di�erences make a di�erence: Individual dissimilarity and group hetero- [35] Richard L Moreland, JM Levine, and ML Wingert. 2013. Creating the Ideal Group: geneity as correlates of recruitment, promotions, and turnover. Journal of Applied Composition E�ects at Work. Taylor and Francis. 644 pages. Psychology 76, 5 (1991), 675–689. [36] Ann-Marie Nazzaro and Joyce Strazzabosco. 2009. Group dynamics and team [20] David W. Johnson and Roger T. Johnson. 2004. Assessing students in groups : building. DEVELOPMENT 4 (2009). promoting group responsibility and individual accountability. Corwin Press. 206 [37] Chinasa Odo. 2019. Group Formation for Online Collaborative Learning: A pages. Systematic Literature Review. In International Conference on Arti�cial Intelligence [21] David W Johnson, Roger T Johnson, Karl A Smith, and Karl Smith. 2013. Co- in Education. operative Learning: Improving University Instruction By Basing Practice On [38] Chinasa Odo, Judith Mastho�, and Nigel Beacham. N.D.. Heterogeneous Group Validated Theory. Journal on Excellence in Universityn Teaching (2013). Formation: Combining the right Learner Characteristics for Online Collaborative [22] M. KankaraÅą. 2017. Personality matters. Organisation for Economic Co-Operation Learning. (N.D.). Unpublished manuscript. and Development (OECD) (2017). [39] Chinasa Odo, Judith Mastho�, Nigel Beacham, and Manal Alhathli. 2018. A�ective [23] Elena Kardanova and Alina Ivanova. 2014. Heterogeneity of the Educational State for Learning Activities Selection. In Proceedings of Intelligent Mentoring System: an Introduction. (2014). Systems Workshop Associated with the 19th International Conference on Arti�cial [24] Saul M. Kassin. 2004. Psychology. Pearson/Prentice Hall. 826 pages. Intelligence in Education, AIED 2018. 1–10. [25] Marjan Laal. 2011. Lifelong Learning: What does it Mean? Procedia - Social and [40] Rena M. Pallo� and Keith Pratt. 2005. Learning Together in Community: Col- Behavioral Sciences 28 (jan 2011), 470–474. laboration Online. 20th Annual Conference on Distance Teaching and Learning [26] Emily R Lai. 2011. Collaboration: A Literature Review Research Report. , 49 pages. (2005), 1–5. [27] Lev S. and Vygotsky. 1962. Thought and Language. The Journal of Mind and [41] J. Piaget. 1936. The origins of intelligence in the child. Routledge & Kegan Paul, Behavior Wintet 8, 1 (1962), 175–178. London. [28] Vivian Miu Chi Lun and Michael Harris Bond. 2006. Achieving relationship [42] Tim S. Roberts. 2004. Online collaborative learning : theory and practice. Informa- harmony in groups and its consequence for group performance. Asian Journal of tion Science Pub. 321 pages. Social Psychology 9, 3 (2006), 195–202. https://doi.org/10.1111/j.1467-839X.2006. [43] Muzafer Sherif and Carolyn W. Sherif. 1965. The Adolescent in His Group in Its 00197.x Setting: Theoretical Approach and Methodology Required. [29] Ioanna Lykourentzou, Angeliki Antoniou, Yannick Naudet, and Steven P. Dow. [44] Kirsten A. Smith, Matt Dennis, Judith Mastho�, and Nava Tintarev. 2019. A 2016. Personality Matters: Balancing for Personality Types Leads to Better Out- methodology for creating and validating psychological stories for conveying comes for Crowd Teams. In Proceedings of the 19th ACM Conference on Computer- and measuring psychological traits. User Modeling and User-Adapted Interaction Supported Cooperative Work & Social Computing (CSCW ’16). ACM, New York, (2019). NY, USA, 260–273. [45] Bruce W. Tuckman. 1964. Personality Structure, Group Composition, and Group [30] Judith Mastho�. 2005. The Pursuit of Satisfaction: A�ective State in Group Functioning. Sociometry 27, 4 (dec 1964), 469. Recommender Systems. International Conference on User Modeling 35, 38 (2005), [46] Rudina Vrioni. 2011. E�ects of Group Learning on the Academic Performance of 297–306. University Students. problems of education in the 21st century 33 (2011). [31] Robert R. McCrae and Oliver P. John. 1992. An Introduction to the Five-Factor [47] Vygotsky. 1978. Integration between learning and development. Harvard University Model and Its Applications. Journal of Personality 60, 2 (jun 1992), 175–215. press, Cambridge. [32] Sinéad McGivney, Alan F. Smeaton, and Hyowon Lee. 2009. The e�ect of per- [48] Mona Zamani. 2016. Cooperative learning: Homogeneous and heterogeneous sonality on collaborative task performance and interaction. Lecture Notes of grouping of Iranian EFL learners in a writing context. Cogent Education 3, 1 (feb the Institute for Computer Sciences, Social-Informatics and Telecommunications 2016). Engineering 10 LNICST (2009), 499–511. [49] Vicky Zygouris-Coe. 2012. Collaborative Learning in an Online Teacher Educa- [33] Catherine Giggs Michael A. Campion, Gina J. Medsker. 2001. Rela- tion Course: Leasons Learned. ICICTE Proceedings (2012). tions_Between_Work.pdf. , 46 pages.