An artificial intelligence-driven learning analytics method to examine the collaborative problem-solving process from the complex adaptive systems perspective

Collaborative problem solving (CPS) requires student groups to solve complex and ill-structured problems without fixed answers in order to achieve the goal of collective knowledge co-construction (Barron, 2000; Roschelle & Teasley, 1995; Vrzakova et al., 2020). In the past decade, CPS has been widely conducted to promote students’ active learning and meaning-making through various learning activities in order to enable students to connect and construct knowledge in a structured manner and create collective knowledge artifacts (Farrokhnia et al., 2019; Wang et al., 2017). CPS involves multiple levels of dynamic interactions between (1) individual students, (2) individual students and the student group(s), (3) individual students and the learning environment or a knowledge artifact, and (4) the student group(s) and the learning environment or knowledge artifact(s) (Stahl & Hakkarainen, 2021). This multilevel and multilayered nature of CPS indicates a certain level of complexity (Damşa, 2014; Stahl, 2013; Stahl & Hakkarainen, 2021).

Complex adaptive systems theory, originating from the field of biology, emphasizes that a system is a complex emergent entity that arises from the interactions between microscopic entities (Byrne & Callaghan, 2014; Holland, 1996; Lansing, 2003). When this concept is applied to the field of education, the teaching and learning processes can be viewed as a complex, self-organizing system, which emerges from interdependent interactions between individuals, groups, and teaching and learning components (Amon et al., 2019; Jacobson et al., 2016; Mitchell, 2009) (see Fig. 1). Furthermore, the CPS process is a complex phenomenon where a group of students constantly coordinates interactive, cognitive, regulative, behavioural, and social-emotional aspects through conversations and actions to adapt to the complex and dynamically changing collaborative context, eventually forming a self-organizing system. Therefore, the complex system perspective breaks the traditional educational research paradigms such as causal models or linear predictability and highlights the use of organic, nonlinear, and holistic approaches to understanding the nonlinear, dynamic evolution of CSCL (Amon et al., 2019; Holland, 1996).

The conception of complexity is not new to empirical research in the CSCL community. For example, Zuiker et al. (2016) combined complex adaptive systems theory as well as situated cognition theory to understand and explain collaborative learning as a system-level social activity. Vogler et al. (2017) used complex adaptive systems theory to explore how a message is shared by students in synchronous online discussions such that it triggered students’ meaningful engagement in knowledge co-construction. Amon et al. (2019) investigated the system-level team dynamics in collaborative programming from a complex systems theory perspective in order to describe the dynamic, multimodal, and complex nature of collaboration. From the perspective of complex adaptive systems theory, these studies started to examine and illuminate how, in the dynamic process of CSCL, collaborative meaning-making processes are interwoven with the multimodality and dynamics of individuals, groups, learning environments, technologies, and resources during the CSCL process. Moreover, the CSCL community has paid attention to analyzing and understanding CSCL processes from the group or system perspectives, such as collaborative cognitive load, group-level metacognition, or group cognition (Dindar et al., 2020; Kuhn et al., 2020; Zheng et al., 2021). Overall, complex adaptive systems theory has the potential to contribute new perspectives to our understanding of educational phenomena, particularly the collaborative learning process, from a holistic, systematic, and non-linear perspective (Amon et al., 2019; Holland, 1996; Vogler et al., 2017).

Grounded in the complex adaptive systems theory, group members in the CPS process form an adaptive, self-organizing system that has a multi-dimensional nature, synergistic relations, and a dynamic evolutionary character (Amon et al., 2019; Barron, 2000; Curşeu et al., 2020; Jacobson et al., 2016; Mitchell, 2009; Stahl & Hakkarainen, 2021). CPS’s multimodal characteristics are reflected in the interactive, cognitive, regulative, behavioural, and socio-emotional aspects of the learning process (Fiore et al., 2010; Kwon et al., 2014; Malmberg et al., 2017; Zemel & Koschmann, 2013). On the interactive dimension, the multimodal interaction in CPS is built through students’ discourse interaction, text interaction, and behaviour interaction, which together comprise the foundation of collaboration (Ouyang & Xu, 2022; Zemel & Koschmann, 2013). On the cognitive dimension, to solve ill-structured tasks, students need to share information and propose ideas to co-construct knowledge with group members (Barron, 2000; Ouyang & Chang, 2019; Vrzakova et al., 2020). On the regulative dimension, students need to understand the task, negotiate and plan the goal of the group, and monitor and reflect on their collaborative progress (Malmberg et al., 2017). On the behavioural dimension, students solve problems through online activities and coordinate their creation of knowledge artifacts as they regulate each other’s behaviour (Fiore et al., 2010; Stahl, 2017). On the socio-emotional dimension, students need to engage in active listening, encourage participation, and contribute towards the creation of cohesive groups in order to foster active engagement, relaxation of tension, and the emergence of social motivation (Kwon et al., 2014; Rogat & Adams-Wiggins, 2015). More importantly, during the CPS process, this multiplicity of dimensions come together to form a complex, interdependent set of relationships that ultimately contribute to an adaptive, self-organizing system with multilevel and multilayered characteristics (Byrne & Callaghan, 2014; Hilpert & Marchand, 2018).

Furthermore, CPS is a dynamic process, where the relations between elements change over time, and ultimately affect the quality of collaboration (Holland, 1996; Hoppe et al., 2021; Koopmans & Stamovlasis, 2016; Saqr & López-Pernas, 2022). During the CPS process, students are required to adjust and coordinate interaction, behaviour, cognition, and emotional contributions to the evolving task context. As they do, they strive to achieve high-quality collaboration reflected in their knowledge construction process as well as the resulting knowledge artifacts (Amon et al., 2019; Barron, 2000; Wiltshire et al., 2019). For example, Amon et al. (2019) found that the regularity of coordinative patterns in students’ speed of communication, body movement, and screen interaction over time in pair programming can influence their collaborative learning outcomes. Saqr et al. (2021) applied various methods (e.g., process mining, sequence mining map, and temporal networks) to investigate the relational and temporal dynamics of students’ self-regulated learning behaviours during online collaborative academic writing. Curşeu et al. (2020) proposed a multi-level dynamic model to examine how the variations of core self-evaluations, study engagement, group development, and relational conflict influences group identification in collaborative learning. In summary, there is emerging evidence that CPS is a multimodal, dynamic, and synergistic phenomenon, which might be better interpreted through the lens of the complex adaptive systems theory. However, studying CPS through this lens poses significant methodological challenges to traditional statistics, which are typically applied to homogenous, unimodal data sources that are frequently used in CSCL research.

Collaborative learning analytics (CLA), an integration of CSCL and LA, is used to explain, diagnose, and promote collaborative learning processes (Wise et al., 2021). Due to the complexity of CPS, merely focusing on one dimension or perspective of collaboration may cause inconclusive and incomprehensible results. Therefore, CLA advocates the possibility to collect complex and multimodal data (e.g., behavioural, physiological, representational data) to develop comprehensive computational, algorithmic models of collaboration (Blikstein, 2013; Borge & Mercier, 2019; Mu et al., 2020; Ouyang et al., 2022; Sullivan & Keith, 2019). Echoing this trend, recent CSCL research has started utilizing multimodal learning analytics (MMLA), integrating multimodal data collection, learning analytics, and AI algorithm-enabled modelling to analyze and mine the complex, dynamic characteristics of CSCL processes (Gorman et al., 2020; Khan, 2017; Olsen et al., 2020; Sharma & Giannakos, 2020; Wiltshire et al., 2019). For example, Khan (2017) proposed a hierarchical computational approach to analyze multimodal data (including audio, video, and activity log files) and modelled the temporal dynamics of student behaviour patterns in collaboration. Wiltshire et al. (2019) used growth curve modelling to examine how students’ multiscale movement coordination (e.g., speech, gestures, mouse and keyboard movement) changed across the duration of CPS. Gorman et al. (2020) used computational, quantitative models (including discrete recurrence, non-linear prediction algorithms, and average mutual information) to detect real-time changes in group communication reorganization patterns during collaborative training. The main premise of this emerging research stream is that, compared to traditional learning analytics, such as social network analysis (e.g., Ouyang, 2021) or content analysis (e.g., Jeong, 2013), integration of algorithm-enabled methods in learning analytics and educational data mining can better deal with complex, nonlinear information, as well as extract and represent multi-level and high-dimensional features of CSCL (de Carvalho & Zárate, 2020; Ouyang et al., 2023).

In this research, we present the results of our investigation of groups’ collaborative patterns during CPS in an online, synchronous collaboration platform. Taking a complex adaptive systems theory perspective to interpret CPS, we collect and analyze the multimodal process and performance data of student groups, including verbal audios, computer screen recordings, and concept map data. Furthermore, we propose a three-layered analytical framework that integrated learning analytics and AI algorithm-driven methods to examine the collaborative patterns and the regularity of those patterns. We aim to answer two main research questions:

The research context was two graduate-level seminar courses, titled Distance and Online Education and Online Learning Analytics offered by the Educational Technology (ET) program at a top research-intensive university in China. The course was designed and facilitated by the same instructor (the first author). She designed a series of CPS activities for small groups (3 or 4 students per group) to work on (see Table 1). Groups were asked to complete one open-ended, ill-structured problem related to the course content in an online, synchronous collaboration platform named huiyizhuo (see Fig. 2). The current research dataset consisted of 24 datasets; each dataset included the group’s audio recordings of verbal communication data (about 2,160 min in total), computer screen recordings of click stream data (about 2,160 min in total), text-based chatting, and the final product of concept map data. The instructor occasionally engaged in the CPS activities and her engagement was not the focus of this research; therefore, to maintain data consistency, the instructor data were all removed.

The instructional design of CPS activities followed the problem-based learning cycle (Hmelo-silver, 2004): students first analyzed the problem scenario, then identified the knowledge gap needed for solving the problem, and finally generated possible solutions through collaboratively building a concept map. Problems were all open-ended, ill-structured problems that did not have a fixed solution. For example, one CPS activity asked the groups to work as a teaching team to design online learning resources for high school classes. The online collaboration platform Huiyizhuo (https://​www.​huiyizhuo.​com/​) was used, which provides functions of text chatting, audio and video communication, concept map, note and comment, and resource sharing. In the CPS process, members first communicated through audio and text chatting, then shared resources and kept communications to share knowledge, and finally constructed a concept map to solve problems and propose solutions (see Fig. 2). The concept map was used as the main tool for the collaborative problem-solving process (Novak & Cañas, 2008).

We proposed and used a three-layered analytic framework to examine the characteristics of collaborative patterns during CPS activities (see Fig. 3). AI algorithms were incorporated into that framework using multiple learning analytic approaches (e.g., quantitative content analysis, clickstream analysis, epistemic network analysis). In Layer 1 of data pre-processing and analysis, we coded the interactive, cognitive, behavioural, regulative, and socio-emotional dimensions of CPS (see Table 2). In Layer 2, featuring a multichannel sequence analysis (MCSA), we examined the similarity of groups’ CPS sequences in order to detect distinct clusters of collaboration patterns. In Layer 3, we used a multi-method approach (including statistical analysis, epistemic network analysis, and hidden Markov modeling) to further examine the regularities among different types of collaboration pattern clusters associated with particular student groups.

ANOVA with the Bonferroni correction was conducted to test for significant differences between the three types on the five dimensions. Before ANOVAs, Levene tests were conducted; and the results confirmed the homogeneity of variance. Moreover, a series of post-hoc pairwise comparisons were conducted to further reveal significant differences between types (see Table 4). Considering that some data (e.g., TU, RM, OB, and EPI) were not normally distributed, a non-parametric test was conducted to cross-check the ANOVA results. The results showed that there were significant differences in the frequency of Int-B, KS, KM, TU, RM, CM, and OB (p < .05) with a Bonferroni correction using the Kruskal-Wallis test. Specifically, there were statistically significant differences between the three types along the behavioural dimension (RM, CM, OB). Regarding the main behaviours of concept mapping (i.e., CM), Type 1 ranked first among the three types, followed by Type 2 and Type 3. CM indirectly reflected students’ cognitive engagement (i.e., KS, KM, and KD) when creating and editing nodes in the concept maps (Type 1: KS: Mean = 9.40, KM: Mean = 18.80, KD: Mean = 61.40; Type 2: KS: Mean = 7.64, KM: Mean = 16.07, KD: Mean = 51.14; Type 3: KS: Mean = 4.00, KM: Mean = 9.60, KD: Mean = 17.20). In Type 3, more OB behaviors (Mean = 146.8, SD = 53.38) were detected than interactive behaviors (RM: Mean = 20.6, SD = 13.70; CM: Mean = 42.2, SD = 7.19). Moreover, OB, to some extent, also reflected students’ monitoring of others’ operations in the regulative dimension. Moreover, on the interactive dimension, statistical significance was found on Int-B, where Type 1 had the highest frequency, followed by Type 2 and Type 3; but no significant difference was found on Int-C, where all three types had a high level of Int-C. On the cognitive dimension, significant differences were found on KS (Type 2 > Type 1) and KM (Type 3 > Type 1, Type 3 > Type 2), except KD (all three types had a low level of KD). On the regulative dimension, significant differences were found on TU (Type 1 > Type 2, Type 1 > Type 3), while no significant difference was found on GSP and MR. In addition, no significant difference was found in the socio-emotional dimension, possibly due to sparsity. We thus concluded that in general there were different collaborative patterns among the three types.

Different characteristics were identified among the three types of collaboration pattern clusters, reflected by the respective locations of the centroid in each epistemic network (shown as red circles in Fig. 6). For Type 1, the centroid of the epistemic network was located at the upper left corner, focusing on the behaviour-related codes (including OB, CM, and Int-B). The connection between Int-B and OB was 0.74; the connection between Int-B and CM was 0.72; and the connection between OB and CM was 0.69. For Type 2, the centroid was located at the bottom of the epistemic network, focusing on the behaviour and communication-related codes (including CM, KM, KS, Int-B, and Int-C). The connection between Int-B and Int-C was 0.71; the connection between Int-C and CM was 0.67; the connection between Int-C and KM was 0.62; the connection between Int-C and KS was 0.61; and the connection between Int-B and CM was 0.60. For Type 3, the centroid of the epistemic network was located on the right side, mainly focusing on communication-related codes (including KM, GSP, and Int-C). The connection between Int-C and OB was 0.90; the connection between Int-C and KM was 0.73; and the connection between Int-C and GSP was 0.59. In summary, Type 1 concentrated on the behaviour-related codes; Type 2 focused on both communication-related and behaviour-related codes, and Type 3 concentrated on the communication-related codes.

From a transitional perspective, similar characteristics among the three types of collaboration pattern clusters were observed in the HMM results. Specifically, a 5-state HMM of Type 1, a 7-state HMM of Type 2, and a 5-state HMM of Type 3 were selected based on fit, where the lowest values for BIC are preferred (see Table 5; Fig. 7, and Fig. 8). Among the HMM results, a regular transition pattern was observed such that communication (State 1 in Type 1, 2, 3), moved to behaviour (State 2 in Type 1, 3 and State 2, 4 in Type 2), and finally returned to communication (State 3, 4 in Type 1, 3 and State 5, 6 in Type 2) (see Table 5; Fig. 7).

A regularity was detected in the transition patterns such that where Type 1 was characterized as a behaviour-oriented transition, Type 2 was characterized as a communication-behaviour-synergistic transition, and Type 3 was characterized as a communication-oriented transition. In Type 1, groups were more likely to transition from other states to the behaviour-related states (State 2, 4) (see Fig. 7). Additionally, states related to peer communications (State 1, 3) were short and infrequent, while states of peer behaviours (State 2, 4) were long and frequent (see Fig. 8). In Type 2, students had a high probability of transitioning to State 4 and State 6, which included both communication-related and behaviour-related codes. Except for State 3 and State 4, each state in Type 2 shared a balance of time and frequency across the CPS processes. In Type 3, groups had the highest probability of transitioning to a communication-related state (State 3). States dominated by peer communications (State 1, 3) were long and frequent, while states dominated by peer behaviours (State 2, 4) were short and infrequent.

Due to the complexity of CPS (Amon et al., 2019; Jacobson et al., 2016; Stahl & Hakkarainen, 2021), unimodal data and traditional statistics may be limited for holistically modelling it (Ouyang et al., 2022). This research collects multimodal data, including speech, screen-capture recordings, and concept map data, and proposes a three-layered analytical framework integrating learning analytics and AI algorithm-driven methods to investigate the collaboration patterns of groups working through an online collaboration platform. Compared to previous studies, this approach proposes a multilayered approach to better understand the complex nature of CPS from an organic, nonlinear, and holistic perspective that is aligned with the complex adaptive systems theory (Vogler et al., 2017; Zuiker et al., 2016). To answer the research questions, three types of collaborative patterns were detected with different levels of the final concept map performance: Type 1 was characterized as a behaviour-oriented collaborative pattern and associated with medium-level performance, Type 2 was characterized as a communication-behaviour-synergistic collaborative pattern and associated with high-level performance, and Type 3 was characterized as a communication-oriented collaborative pattern and associated with a low-level performance. More importantly, the attributes of the three types were together demonstrated through the quantitative frequency, structural, and transitional aspects as the results showed. Regarding the behaviour of concept mapping (i.e., CM), Type 1 ranked first among the three types, followed by Type 2, and Type 3. Since the complexity of CPS, the cognitive dimension was also reflected through students’ online behaviours as knowledge sharing and construction during the concept mapping process, which partially explained the relations between the groups’ behaviours and final performances. For example, Type 3 had the lowest level of cognitive engagement (KS, KM, and KD) through online behaviours, which also resulted in the lowest score of the final performance among the three types. The ENA and HMM results also indicated different characteristics of the three types. For example, Type 2 shared strong connections between communication-related and behaviour-related codes with a communication-behaviour-synergistic transition; Type 1 generated strong connections only within the behaviour-related codes with a behaviour-oriented transition; Type 3 generated strong connections only within the communication-related codes with a communication-oriented transition. Overall, this research revealed three types of collaborative patterns in CPS as well as their characteristics to further verify the multimodality, dynamics, and synergy of the CPS process.

The complexity theory, as an emerging paradigm in educational research (Morrison, 2002), looks at and examines the educational system in ways that break with the simple cause-and-effect models, linear modelling, and regression and replaces them with organic, non-linear, and holistic approaches (Cohen et al., 2013). The recently published international handbook of computer-supported collaborative learning calls for the application of AI algorithm-based models to model complex systems and learner self-organization for collaboration at various scales (Cress et al., 2021). From the perspective of complex adaptive systems theory, CPS is a complex phenomenon, in which group members interact with each other, the learning environment, or the knowledge artifacts adaptively to form a group-level, collaborative pattern and collective intelligence (Jacobson et al., 2016). This research provides a new method to explain the emergence of a self-organizing system during the CPS process by integrating AI algorithms with learning analytics and various modalities of data to analyze multimodal, dynamic, and synergistic characteristics of collaboration.

First, the results demonstrated the multimodal characteristics of collaborative learning, reflected by the interactive, cognitive, regulative, behavioural, and socio-emotional dimensions, which formed the foundation for the emergence of an adaptive, self-organizing system as observed within Markov models’ distinctive state-transition patterns associated with different types. Consistent with previous studies (e.g., Ouyang & Chang, 2019; Park et al., 2015), we found that groups’ peer interactions were primarily reflected in students’ verbal communications, and the cognitive and regulative dimensions were closely related to peer interactions via verbal discourses. In other words, active peer communications formed a foundation for deep-level knowledge construction and group regulation (Zemel & Koschmann, 2013). On the behavioural dimension, collaborative patterns with high and medium performance had more active behaviours (i.e., concept mapping, and resource management), while low performing patterns were associated with more passive behaviours (i.e., observation). In addition, the cognitive dimension was also reflected in students’ online behaviours as knowledge construction in concept mapping, which might partially explain the relations between groups’ behaviours and final performance. For example, our results showed that Type 3 had the lowest level of knowledge construction through students’ online behaviours, which also resulted in the lowest collaborative performance among the three types. Moreover, more deep-level knowledge was generated through concept mapping behaviours than verbal communications. Therefore, in the multimodal CPS process, we are able to infer that online behaviour might play a critical role in transforming students’ superficial-level and medium-level knowledge in verbal discourse into a deep-level knowledge co-construction (Ouyang & Xu, 2022). However, inconsistent with previous studies that highlighted the role of social emotions (e.g., Kwon et al., 2014; Rogat & Adams-Wiggins, 2015), no significant difference was found between collaborative patterns on the social-emotional dimension. Apart from that, the socio-emotional dimension was weakly-connected with other dimensions. One of the potential reasons is that there may be a cultural aversion to students expressing their social emotions during collaboration (Zhang, 2007, 2013). These results indicate the complex connections between multiple dimensions emerging into the multimodality of CPS, which in turn influences the collaborative outcomes of group interactions.

Second, the dynamic characteristics uncovered how an adaptive, self-organizing systems emerged within groups during the CPS processes as observed in Markov models’ distinct states for each type of pattern. When confronted with a new problem, groups tended to discuss it together to understand the task and set a specific goal (i.e., communication state). Initial verbal communications in this stage helped groups of students adapt to the social-collaborative environment and become familiar with their group members, which is beneficial to transitioning students into active collaborators (Kwon et al., 2014). However, less cognitive engagement was found in the initial communication stage. Since meaningful knowledge construction belongs to a relatively higher level in CPS, it might require more complex factors (Roschelle & Teasley, 1995). It might also appear based on enough peer interaction and the feeling of safety and collaboration (Schindler & Bakker, 2020; Zemel & Koschmann, 2013). After that, some students started operating concept maps or managing resources, and others kept observing (i.e., behaviour state). In this stage, students’ cognitive engagement and metacognitive regulation are mainly reflected through their online behaviours and sometimes presented by means of verbal communication. When new difficulties or bottlenecks appeared, students would suspend their work in order to make time to discuss it again and reflect on how to solve the problems (i.e., communication state). This research demonstrated how groups evolved through the developmental stages (i.e., from communication to behaviour and back to communication) to adapt to the ongoing changes in collaboration demands and finally form a self-organizing system with relatively stable states and their associated transition probabilities.

Third, as one of the keys to succeed in forming an adaptive, self-organizing system during collaboration, the synergistic characteristics were embodied in groups’ coordination between communication and behaviour. Similar to previous studies (e.g., Amon et al., 2019; Ramenzoni et al., 2012; Wiltshire et al., 2019), these findings indicated that high-performing groups (i.e., Type 2) gave rise to synergistic collaboration across verbal discourses and online behaviours in order to adapt to the shifting task demands. However, considering the external factors (e.g., task difficulty, group configuration, social learning context), it is challenging for all groups to collaborate synergistically and to achieve a high quality of collaboration. For example, the medium-performing groups (i.e., Type 1) were more likely to involve online behaviours (e.g., concept mapping) to drive CPS activities, while the low-performing groups (i.e., Type 3) were more likely to conduct CPS activities through verbal discourses. Furthermore, the ENA connection values were relatively stronger in Type 1 and Type 3 than in Type 2, which to some extent verified the difficulty of achieving a high-quality synergy between communication and behaviour in the CPS processes. Overall, this research highlights the complex characteristics of collaborative learning, which further argue for multi-dimensional, multi-level, and multi-perspective analytical approaches to collaborative learning research (Cohen et al., 2013; Jacobson et al., 2016; Morrison, 2002). This research builds a bridge between the complex adaptive systems theory and analytics of collaborative learning to guide future empirical studies in the CSCL field that would take similar approaches to the investigation of students’ group interactions.

Since collaborative learning is a complex, adaptive process, this research extends investigations on using AI-driven learning analytics to understand multimodal, dynamic, and synergistic characteristics of groups’ collaborative patterns during a complex collaborative task. On the one hand, multimodal data collection and analysis are suggested for the investigation of complex educational phenomena and problems (Cukurova et al., 2020; Michail Giannakos, Roberto Martinez-Maldonado; Stahl & Hakkarainen, 2021; Vogler et al., 2017). Recent research has used multimodal data (e.g., speech rate, gesture, body movement, eye movement) to examine the collaboration patterns and characteristics in CSCL, which can complement analysis results from traditional discourse, online content, and product data (Amon et al., 2019; Blikstein, 2013; Mu et al., 2020). More importantly, due to the complexity of CPS, the multiple dimensions (e.g., interactive, cognitive, regulative) might be reflected through communication and the behaviour of collaboration. For example, the behaviours in this research also indirectly reflected the cognitive engagement and metacognitive regulation of the groups of students during the CPS processes. On the other hand, this research verified that the AI algorithm-enabled learning analytics and data mining approach can help to reveal complex characteristics of CPS from quantitative, structural, and transitional perspectives. Compared to traditional learning analytics methods, the integration of learning analytics approaches and algorithm-enabled methods can better process multimodal and nonlinear data in order to extract and represent the complex and dynamic structure of CSCL (de Carvalho & Zárate, 2020). For example, HMMs have the potential to capture the stable, distinct states of CPS as well as dynamic movements between and within them, which has the particular ability to reveal the adaptive, self-organizing character of collaborative learning. Furthermore, advanced AI algorithms (e.g., multidimensional recurrence quantification analysis, natural language processing, genetic programming) have been applied in CSCL to analyze and reveal the complexity and dynamics of collaboration (Amon et al., 2019; de Carvalho & Zárate, 2020; Hoppe et al., 2021; Sullivan & Keith, 2019). Future work can further integrate these advanced AI algorithms with learning analytics and data mining to reveal the multilevel, multidimensional characteristics of CSCL. There is a call for the application of technological advances (such as machine learning) to foster real-time feedback and assessment of group work, identify collaboration patterns to effectively guide productive collaboration, and accurately predict collaborative outcomes (Amon et al., 2019; Eloy et al., 2019; Xu & Ouyang, 2022).

CPS is a complex process that is unlikely to follow a particular model of linear sequences of actions. Therefore, during the CPS process, instructors should provide dynamic scaffolding as well as adaptive support to promote students’ collaboration. First, instructors can provide different scaffolding to enhance groups’ collaboration quality based on the multimodal characteristics of CPS. For example, the foundational role of active communications was reinforced by the research results, therefore, instructors are able to regulate group communications through metacognitive scaffoldings (e.g., monitor and stimulate student tasks) to foster students’ cognitive and regulative engagement (Ouyang et al., 2021; Van Leeuwen et al., 2015). In addition, due to the weak connection with the socio-emotional dimension as indicated above, the instructor’s social support, such as encouragement and affirmation, might ease the atmosphere and mobilize students’ social emotions (Ouyang & Scharber, 2017; Ouyang et al., 2020). Second, instructors are supposed to provide flexible and changeable support (Park et al., 2015; Van de Pol et al., 2010) in terms of student groups’ dynamic evolvement in CPS (i.e., from communication to behaviour and back to communication) we detected in the current study. To be specific, in the initial stage, instructors can regulate students to organize discussion about the topic; in the middle stage of online operation, instructors can reduce intervention, observe students, and provide specific support according to the particular needs of the group; in the latter stage, instructors can guide students to solve problems through information relevant to the topic (Kaendler et al., 2015; Ouyang & Xu, 2022; Van Leeuwen et al., 2015). Third, it is challenging to achieve the high-quality synergy between communication and behaviours of student groups. Thus, instructors should pay attention to groups’ inactive moments in verbal communication or online behaviour to foster synergistic meaning co-construction (Barron, 2000; Brown et al., 1989; O’Donnell & Hmelo-Silver, 2013). When students only focus on the discussion and neglect operating the task, the instructor can remind them to return to online operations to drive the task; when students lack enough communication, instructors can guide them to discuss the topics together. Taken together, from a pedagogical perspective, instructors should be aware of groups’ dynamic and synergistic status, and support their work appropriately with varied scaffoldings to promote students’ CPS.

There were four major limitations in the current research. The first limitation of this research is the sample size of student groups (4 groups generated 24 collaborative activities), with a limited range of demographic backgrounds. The small sample size might be a reason for causing the insignificant differences in the final concept map products. Therefore, future empirical research needs to expand the sample size and experiment with different courses to test, validate, or modify the implications. The second limitation is that we do not control the instructor’s participation during the CPS processes. The instructor occasionally participated in the collaboration to provide social, cognitive or regulatory support, which might influence students’ collaborative patterns. Due to the authentic and natural instructional context, we cannot control all external factors in the CPS processes; in the future, a quasi-experimental approach can be considered to compare collaborative patterns occurring in two contrasting conditions. The third limitation is that we only collected students’ communication discourse and online behaviours to analyze the CPS processes. There are complex, interrelated connections between dimensions. For example, a student’s operation on a concept map is coded as the behavioural dimension in the current study, which also involved cognitive and regulative dimensions (e.g., searching for information). Detailed analytics are needed to further uncover complex interrelationships. Moreover, due to the complexity of CPS, other modalities of data (e.g., physiological, and eye tracking data) might provide further insights into CSCL processes. Finally, regarding the AI algorithms, we only used HMMs to detect the temporal patterns in the collaborative process, which might not be the most efficient and accurate algorithm-enabled method. However, this is a significant initial attempt, considering that HMMs have rarely been applied to multichannel social sequence data, especially in the field of education. Future work is encouraged to apply various AI algorithms to compare the relative efficiency and accuracy they might achieve. In addition, other machine learning techniques can be used to guide real-time feedback and predictions of collaboration patterns and outcomes.