In the OECD countries, the proportion of young people between 25 and 34 years of age with a university degree rose by a total of 14 percentage points between 2000 and 2013 (OECD, 2020). As one measure to increase number of students, universities have opened up to 'non-traditional students'. The term 'non-traditional students' refers to the socio-demographic diversity of learners compared to traditional participants in higher education (Chung et al., 2014). However, there are different views on the characteristics that can be used to identify 'non-traditional students'. Chung et al. (2014) identified 13 categories of socio-demographic characteristics in a literature study. One part of these categories are based on the person (age, sex, ethnicity), the state of health (disability or trauma), the current living situation (multiple roles) and other sociological factors (being disadvantaged, demographically different from norm). The other part of catagories point out differences in the admission pathway and in the course of study (mode of study, gap in studies, enrolment in non-traditional programs, previous degree, commuter status).

Diversity is equally evident in learners' learning-related abilities and characteristics such as self-efficacy (Menold and Jablokow, 2019; De Clercq et al., 2020; Pluut and Curşeu, 2013), cognitive style diversity (Menold and Jablokow, 2019), learning styles, lexical density and lexical diversity (Gregori-Signes and Clavel-Arroitia, 2015) as well as the digital skills (Hatlevik et al., 2015). These learning-related abilities and characteristics may in turn be determined by sociodemographic characteristics that indicate social milieus. In this context, Bremer (2004) points out, for example, that the ability for self-regulated learning is shaped by the social milieu.

A comprehensive body of empirical work demonstrates that diversity has an impact on learning behavior and thus on learning outcome. For instance, De Clercq et al. (2020) were able to attribute performance differences in the first year of study to differences in the high school grade, socioeconomic status, informed-choice, and self-efficacy beliefs. With regard to collaborative learning, Pluut and Curşeu (2013) showed the influence of gender, nationality, and life experience within a group on the performance in creativity tasks.

The influence and impact of diversity on learning can be substantiated not only empirically, but also theoretically. According to the situated learning theory (Lave and Wenger, 1991) it can be argued that a fit of learning material (e.g. case studies) to a learner's experiential background can increase learning success. A discrepancy between relevant learner characteristics and the design of the learning environment can therefore be detrimental to learning.

From the perspective of teachers, this raises the question of how diversity characteristics can be taken into account in teaching. General recommendation like the use of multicultural language in online courses (Good et al., 2020) or the use of gender-appropriate teaching, contents, or language aims at a uniform implementation for all learners. In most cases, however, teachers are unaware of the diversity characteristics of their students, both individually and aggregated across a cohort of students. This is due to a variety of reasons. First, there is a consensus that individual learners should not be disadvantaged or advantaged on the basis of certain characteristics. It is therefore forbidden, for example, to gather information about socio-economic status or ethnic origin and to derive immediate didactic actions from this. These ethnic and moral boundaries are therefore regulated in laws against discrimination, but also prohibited through laws like the GDPR (European Commission, 2016). Second, there are characteristics such as learning style or self-efficacy, which the learners themselves are often not aware of and which can only be determined using standardized survey instruments.

On the other hand, educational institutions collect a necessary set of sociodemographic characteristics when a student enrolls to a study program (e.g. date of birth, gender, address). Information about the course of study (e.g. prior knowledge from past courses) is available from university IT systems. Both information sources are not available to lecturers. The learning behavior within online learning environments, including LMS, is mostly captured in detailed log files, but is only made accessible to a limited extent using learning Analytics methods like dashboards (cf. Schwendimann et al. (2016)). Learning-related characteristics or interactions are more likely to be grasped accidentally by teachers, for example in conversations (e.g. lexic diversity) or by assessing learning performance.

This creates a dilemma. On the one hand, teachers are confronted with an increasingly diverse student body and have to deal with it. On the other hand, they are only able to grasp learning-related diversity characteristics to a very limited extent and thus cannot take them into account in the design and implementation of learning opportunities. Since many educational institutions have intensified online learning in the context of the COVID-19 pandemic, diversity characteristics need to be considered in blended and online learning, too.

LMS as a prominent representative of online learning environments should support teachers in recognizing diversity and incorporating it into online learning experience. However, in a typical LMS like Moodle, Ilias, Canvas or Blackboard (Ifenthaler, 2012) all learners of a course are usually confronted with the same learning environment, although their learning characteristics and learning interactions are by no means homogeneous. Thus, learners are largely treated the same in terms of learning materials, instructions, assessment options, etc., despite differences in educational background, work experience, personal skills, needs, etc. The diversity of learners is in contradiction to a uniform learning environment, since different preferences and, in the course of learning, different needs may require adaptation of the learning environment in terms of learning materials, instructions, feedback and forms of interaction.

A promising solution here is to personalize the learning environment so that each learner gets a suitable learning environment. Current approaches to personalizing learning environments often use a one-time cofinguration of the learning environment at the beginning, which depends on an analysis of the parameters relevant for the configuration. For this purpose, e.g. questionnaires, diagnostic tests, or data from previous courses can be used to determine the parameter values for each learner. Such a configuration makes it easier for learners to access information, as only the relevant ones are displayed. For example, only the information of the exercise groups to which a learner is assigned is displayed.

Or the material and tasks from knowledge areas in which a learner did not achieve good results in the entrance test are recommended. However, this kind of personalization through initial configuration does not support adaptation to learners' changing needs in shorter periods of time, which may result, for example, from a current drop in motivation, an accumulation of missed deadlines, or sudden problems in a learning group.

Figure 1: Cycle of a diversity-related personalized learning environment that is individually adapted through learning analytics

In such a situation, the goal is to dynamically personalize the learning environment to provide personalized learning support that adapts to learners' current needs. For this purpose, the relevant characteristics, which may vary in the entirety of the learners, are mapped for each learner onto an individual learner model (see Figure 1). On the basis of the current characteristics of the learner model, the personalized learning environment is created or, if necessary, adapted in the event of changes. The learning behavior can change with respect to different characteristics on different time scales. Examples are deviations of the current behavior from a previous work plan, errors in self-testing and exercises, etc. Such feature changes can be captured with learning analytics approaches and lead to an update of the respective learner model: an adaptation loop is created. Because it cannot be assumed that ex-ante appropriate adaptations of the learning environment for all occurring behaviors can be defined, a machine learning approach that can determine the appropriate personalization for the learner in each case based on student data (diversity characteristics, interactions, learning goal achievements) and data about learning opportunities (programs, courses, learning activities) seems necessary.

With the help of such an approach, the implementation of a personalized adaptive learning environment becomes achievable, which provides personalized dashboards, learning materials, interaction opportunities, and support based on a dynamic learner model and tutoring model according to the learner's needs.

A good overview of the current state of research and the different approaches to design and implement adaptive personalized learning environments can be derieved from systematic reviews of literature (Vandewaetere et al., 2011; Grubišić et al., 2015; Xie et al., 2019; Talaghzi et al., 2020; Martin et al., 2020) and systems (Newman et al., 2016; Osadcha et al., 2020). In addition, systematic literature reviews can be found on specific aspects of adaptive learning environments such as the learning style (Akbulut and Cardak, 2012; Kumar et al., 2017) and the user model (Nakic et al., 2015). Grubišić et al. (2015) specifically addresses the requirements for adaptive courseware. Only a few reviews examined the adaptation methods used (Vandewaetere et al., 2011; Talaghzi et al., 2020), while the system architectures within the educational institutions were not included in any of the mentioned survey papers. In the following section we will exemplify two use cases to outline the future of LMSs as one building block in a broader institutional environment of heterogeneous systems and data sources.

The first use case shows how to consider diversity characteristics such as prior knowledge, professional experience and learning progress when providing personalized learning material (see Figure 2). The provision of such an adaptive personalized learning environment requires the LMS to select appropriate learning resources from a database of all learning resources for a given course. Therefore, the LMS maintains a pool of learning resources, such as slides, course texts, tests, videos, book chapters, with associated metadata.

Figure 2: Selecting learning resources and assessments considering a learner model including behavioral data from the LMS as well as enrollment and examination data from other sources.

In order to adaptively personalize the learner's learning environment according to diversity characteristics, the LMS needs to know the respective characteristics of a given learner. For this purpose, the adaptive personalization functionality of the LMS uses data about the learner available from data sources existing in the IT environment of the educational institution (e.g. data about enrollments, examinations, course selections) and interaction in the LMS to construct a learner model for each learner. For example, enrollment data can be used to determine the professional background of a learner (e.g. completed courses of study or vocational training), while data about finished courses and exams may be used to assess prior knowledge, and interaction and test data of an LMS may be used to measure learning progress.

This learner model reflects the learner's diversity characteristics, and is used to create a learning environment suitable to the current needs of the learner. For example, a learner with a professional background in business administration is provided, if available, with case studies and exercises in the business administration domain while another learner with a background in the health care domain may be provided with case studies and exercises in the health care domain. The learner model enables the adaptive LMS to assess prior knowledge on given concepts or competences based on prior course selection, exams and assignment results. If sufficient prior knowledge exists, the learning environment may offer short summaries including references to background material instead of the usual comprehensive course material and exercises. If tests show deficits, additional hints, learning material and tests may be offered, so that the learner may close knowledge gaps. In this way, a simple form of digital learning support may be implemented.

Figure 3: Adaptive LMS architecture for use case 1

In principle, a platform for implementing such adaptive personalized learning environments needs to support the following steps in the adaptation cycle (see Figure 3):

In the literature, many approaches to support online adaptive learning have been described (cf. Talaghzi et al. (2020)). While previous work focuses on providing specific adaptation features regarding content or assessments, our approach addresses the needs, firstly, to embedd such support in an existing IT infratsructure, and secondly, to support an ongoing adaptation loop that allows continuous learning of adaptation rules (i.e., learning on the meta level: learning how to better adapt). The proposed architecture thus provides a framework for integrating other, more specific adaptation approaches and is open to make such adaptation rules available for new LMSs or learning tools, if they provide the required interfaces.

In the second use case we are looking at the participation of group members in a web conference while working on a joint solution in a shared editor, which is a component of the LMS. As shown in Figure 4, three students are jointly solving a problem that was outlined in a LMS course. In parallel, they discuss their work by means of a web conferencing system extending the LMS. The task requires active and lively collaboration from all group members, including joint writing. This collaboration is essential for a successful solution of the task because it requires the integration of different perspectives. Figure 4 shows that two of the group members (shown at the bottom of the screen) work very well and closely together. The third member (shown on the top left of the screen) does not participate well, does not communicate suitably for an efficient group learning - indicated by the distance from the other members - and appears dissatisfied e.g. with his position in the group or with the other group members - indicated by his negative facial expression. His position in the group might lead to poor participation at solving the problem. The person may be excluded from discussion and will not be able to contribute to the group's solution. Consequently, he will not benefit from the learning process as the other group members do.

Figure 4: Participation support in an online conference

In this use case, the collaboration in the learning group is supported by an Educational Chatbot (cf. Perez et al. (2020)). This bot perceives and analyses the described problems in collaboration and communication of the group. Among other things, it considers the diversity characteristics of communication behaviour, emotions, and learning progress. To do this, it not only uses data captured by the web conference system, but also data from the LMS such as individual text contributions and writing progress in the shared editor. The analysis leads to personalised provision of feedback for the group discussion. The Educational Bot takes into account the individual contributions to the conversation (e.g. the number of words), the emotions from voice sentiment analysis, or lexical analysis in order to relate the terms used with the learning resources in the LMS. It then “takes the floor" and thus addresses all group members. It gives them advice on communication, collaboration, and coordination. In this way, it raises group members' awareness and contributes to the recognition and elimination of problems regarding the group process. Its feedback ensures that all group members participate successfully in the learning process. Figure 5 shows in detail the technical architecture behind the use case.

Figure 5: Architecture Use Case 2

In order to generate appropriate feedbacks, existing data is taken from both, the LMS and the web conference system. This is followed by data preparation, in which heterogeneous data is converted into a suitable format that allows automatic further processing by the system. Specific properties are extracted from the LMS data for later statistical analysis and suitably processed – symbolised by a histogram in the left side of the figure. These include, among other things, the time each member speaks and the interconnectedness, i.e. who communicates with whom. The text resulting from the captured audio conversations of the group members are converted into a format that can be interpreted by the system using Natural Language Processing (NLP) methods - symbolised by a text page.

The sentences spoken in the web conference are first transcribed from the audio format into a text format. They are then subjected to the same steps as described by the processing of the LMS data (see above). In addition, a possible basic emotional attitude of group members is inferred from the voice pitch and possibly the analysis of the facial expressions using computer vision algorithms. This is symbolised in figure 5 by the two emoticons.

In the next layer - analoguous to the description of use case 1 - the data provision in form of a data warehouse and the semantic modelling takes place. The model wharehouse includes the following model types: group model, learner model, pedagogical model and the domain model. The latter three models are discussed in section 3.1.3. The group model represents and characterises the group as a whole and offers information related to pedagogical and social aspects and is formed from the individual student models (e.g. Jaques et al. (2002); Vizcaíno et al. (2000)). It takes into account, e.g., degree and type of participation, abilities, type of prefered exercises, motivation, made mistakes of the group members.

The transition from the data preparation layer ("extract, load, transform") to the data provision and model building layers ("data warehouse and model warehouse") is performed continuously during the group's use of the system. Whenever data or models change due to group members' interactions consistency must be maintained by updating dependent data or models.

The services of the data warehouse and model warehouse are made available to the LMS via an Application Programming Interface. Similar to use case 1, in the next layer the adaptation engine uses the models provided by the model warehouse and, if necessary, an adaptation engine adapts the personalized LMS. Data analysis techniques, e.g. clustering and association mining, enable the evaluation and learning of new personalisation rules. In this use case, techniques for SNA are added.

The upmost Data Usage Layer representing the Online Learning System uses the services of the data analysis, adaption and personalization layer via the ARI. The Data Usage Layer includes the conferencing tool and Educational Bot, both connected to the LMS.

The architecture described (see Figure 5) enables the Educational Bot to provide the above mentioned personalized feedback to all group members on communication, collaboration, and coordination. Thus, it supports the awareness of the group members and helps to address problems regarding the group process.

In contrast to support of group participation, Kim et al. (2020) developed a chatbot agent that monitors and promotes discussions among group members. The bot fullfills the task of moderating the group less as a tool, more as an acting member. To do this, it monitors the participation of the members and, if necessary, encourages them to contribute to the discussion, reminds them of the remaining time and promotes the exchange of diverse opinions. Trausan-Matu (2010) deals with the analysis of conversations of students in an online messenger in a CSCL context. NLP and social network analysis (SNA) are used to analyze chat histories for tutors and to create interactive graphs to help them evaluate and assess the collaboration. Ferguson et al. (2013) uses a combination of natural language processing techniques and machine learning methods to develop a self-learning framework that identifies exploratory learning dialogues based on synchronous text chats. It was tested using a two-day online conference.

In Figure 6 a layered architecture of an adaptive personalized learning environment is shown. This architecture is based on access to various data sources in the bottom layer. Such data is preprocessed in the layer above (extract, load, transform (ELT) layer) and persisted in a data warehouse or model warehouse in the third layer for further use. These three lower layers are independent of a specific application and form the basis of adaptation and personalization of a learning environment consisting of different tools, applications and LMS. The two layers on top of it can access the data in the data warehouse via an API to first perform analysis or determine adaptations before the data usage layer finally provides the adaptive personalized learning environments that can be used by learners and teachers.

The LMS plays an important role in this architecture, but not necessarily a central one. LMSs are just one representative of online learning systems that, among other learning and teaching related applications, are available for the stakeholders involved. However, an LMS is not understood as an atomic entity, but as a modular, distributed, and extensible system. For example, the database of an LMS is separated from the actual application program and the LMS extensions (cf. plugins) are interchangable. The processing steps in the upper two layers can be implemented as part of the LMS or in separate service. While current LMS are focused on content delivery and organizing learning in a static way, they are now considered both, a source of interaction data to be used for analytics and adaptations, and a receiver of adaptation requests regarding content, user interface and possible interactions for a specific learner or learning group. The LMSs of today are not obsolete, as long as they provide the necessary interfaces and thus pave the way for iterative further developments. This is the case with LMS like Moodle, Ilias, and StudIP. In the following subsections, the five layers of architecture corresponding to Figure 6 are explained from bottom to top.

Figure 6: Principle architecture supporting adaptive personalized learning in LMS and other online learning systems

In the application layer there are applications of different kinds (e.g. chatbots, recommender systems, online learning systems), which enable personalization and adaptation in different ways and therefore have to rely on the data in the previously described layers below. One of these applications is the classic LMS as an example of an online learning systems. However, the data storage of these applications takes place in the data source layer, so that, for instance, the log data generated in an LMS can be captured and used for analytics and adaptation as well.

Of special importance in this layer is an extension of Online Learning Systems, i.e. of LMSs and other learning applications, which provides an interface for the implementation of adaptations. We refer to this extension as the ARI. In practice, the ARI is used by the adaptation engine to perform adaptations according to the adaptation rules. It maps abstract adaptation actions defined in the rules on to specific application changes. Its main task is to allow the personalization of content, didactics or presentation visible to the learner in the user interface. For this purpose, specific methods of the learning environment (e.g. direct messages, system notifications), browser specific functions (e.g. alerts, browser-notifications) or specially implemented procedures (e.g. text highlighting, change of text order, modal dialogs, hyperlinks) can be used. In order to use these methods in a way that does not interrupt the user in the learning process, context information and also reactions to past adaptations have to be taken into account before an adaptation is made effective.

In order to implement such an architecture, not all components have to be realized from the beginning and not every layer has to be fully implemented. It makes sense to begin with vertical slices, in which, for example, the data of a source is preprocessed and made available for an immediate analysis within an application. Iteratively, data sources can be added, preprocessing can be automated, and analysis and adaptation methods for different applications can be improved and compared. Even if an implementation of the architecture starts small, it is important to keep from the beginning requirements in mind that are crucial for further expansion and scaling of the system.

The architecture already gives an impression of which components a next generation LMS requires and how they interact. However, this schematic view does not take into account some of the challenges that play a role in development, deployment, and also in long-term operation. Particularly noteworthy are the needs of teachers and learners, who so far have had very little experience with personalized learning offerings in general and adaptive systems in particular. In the following subsections, we will address technical requirements regarding the system architecture and the interaction of data sources, models, and applications as well as the requirements of teachers and learners regarding such learning environments.