1 Introduction

Higher education is a demanding environment that poses new challenges for the learning behavior of students. In contrast to school, in which students’ learning is structured and supported in more detail by their teachers (Vosniadou 2020), university students have to actively plan, monitor, and control their learning behaviors largely on their own to achieve their goals. Yet, many students are struggling in this process, potentially explaining severe dropout rates from study programs (see, e.g., Heublein et al. 2022). The resulting negative consequences accumulate among individuals, leading to major challenges also on a societal level (e.g., loss of time and funds).

With the growing role of digital learning environments in higher education, more and more learning data from students is automatically captured that can be leveraged to actively support them in their learning process. For instance, log data can serve to identify students at risk of dropping out of a course (Foster and Siddle 2020), allowing instructors to proactively support them. In a similar vein, digital learning environments offer scholars and practitioners novel opportunities to implement a wide range of behavioral interventions to automatically support students’ online learning. Recent examples are components that provide feedback on performance (Leung et al. 2022), that help students to monitor their learning progress (Yoon et al. 2021), or that illustrate the online learning time spent (Günther 2021).

Despite the potential of behavioral interventions within digital learning environments, existing research in this area has neglected that students’ personality and learning strategies are inherently different. More precisely, self-regulated learning theory (see, e.g., Pintrich 2004) implies that students employ different learning strategies and therefore might need personalized guidance for their learning. However, such personalized interventions require high efforts for human instructors (Hogan und Pressley 1997). Thus, personalized interventions are so far hardly scalable for a wide range of courses.

We argue that digital learning environments can empower such personalized guidance at great scale for university courses using online content: Through combining vast amounts of user activity data with students’ course performance data (from previous runs of a course), the learning platform can identify patterns which learning actions have been influential to master a course. Ultimately, when these patterns are deployed in a digital learning environment, a corresponding feedback component can provide current students with personalized instructions on how to improve their online learning and potentially so their course performance.

Against this backdrop, we developed a feedback component that leverages the potential of digital learning environments, which we present in this paper. Specifically, we will briefly summarize our technical approach to initialize the feedback component, show its feedback design, and shed light on our experimental approach to test its effects on learners. The paper concludes with a brief description of the anticipated contributions of the study and the planned next steps.

2 Research Design

For our feedback component, we have instantiated machine learning (ML) models that learn relationships between students’ digital learning actions and their overall course performance from past runs of a corresponding course. These models, each of them is solely used for feedback provision in a specific week of the course, are embedded into the learning platform. For each course participant, a week-specific ML model predicts a participant’s performance in the final exam of the course, based on the participant’s past behavior (log data, time tracking data) and their characteristics (socio-demographic backgrounds yielded from the registration page). To subsequently provide personalized feedback, we employ counterfactual explanation methods, which are a recent technical innovation in the field of explainable ML. Counterfactual explainers estimate how model input parameters (i.e., features) need to change in order to achieve a desired model outcome. Embedded in our feedback component, the explainer method infers what additional actions a learner has to perform (that is the change in input parameters) to improve their exam performance (that is the ML model’s output), which Fig. 1 illustrates. The feedback component displays the obtained changes in input parameters as actions for exam improvement. By contrast, the ML model’s predicted exam performance and the potential for exam improvement are not displayed (we treat these just as internal technical metrics).

Fig. 1
figure 1

Technical approach

The displayed actions for learners are based on learning strategies and are presented as instructions for learning behavior in the digital learning environment. To put that into context, instructions such as “Watch the video of lecture 3 again” or “Do the quizzes of Lecture 1” should encourage students to catch up with the learning content, monitor their knowledge, or deepen their understanding of specific topics of the lecture. Figure 2 displays the feedback component with examples of personalized actions for learners to improve their performance. The component is embedded into the main course page of the associated digital learning environment (i.e., open edX) so that it is salient to the learners.

Fig. 2
figure 2

Feedback based on counterfactual explanations

To evaluate the effects of the feedback component, we are running two experimental studies: One in a bachelor level course (summer semester 2022) and one in a master level course (winter semester 2022/2023). Each study follows a difference-in-differences experimental design. More precisely, we provide a group of students with feedback after a baseline phase, while the control group does not receive any feedback. We hypothesize that our feedback component will have desirable effects on students’ course success (for an overview of learning techniques see Dunlosky et al. 2013 and for the effectiveness of feedback see Hattie und Timperley 2007). We investigate course success in terms of exam taking rate and points in the exam. To better understand the effects of the feedback intervention, we additionally conduct a pre- and post-survey that capture influential psychological constructs on online learning. These constructs relate for example to students’ self-regulated learning skills, their procrastination behavior, and feedback acceptance. Additional analyses using these constructs will allow us to understand which students implement the instructions from the feedback component and benefit most from them in terms of study behavior, exam participation, and exam grade. Figure  3 shows the experimental setup for each of the two studies.

Fig. 3
figure 3

Experimental setup

3 Conclusion

In this paper, we have presented a novel component that applies counterfactual explanations for providing personalized feedback. Even though we still have to statistically evaluate the effects of this feedback component on learners, its underlying technical approach is promising. The approach has the potential to unite learner characteristics (e.g., self-regulation skills, susceptibility to procrastination, etc.) and behavioral data to derive personalized guidance for learners on how to improve educational success. In doing so, the feedback component, operating in a digital sphere, can potentially promise more benefits than those that can be expected from a human instructor: Providing personalized feedback at scale. Given the relevance of learning strategies on academic success in higher education (Broadbent und Poon 2015), this paper encourages practitioners and scholars to consider such scalable approaches for empowering personalized learning support.