Explainable Artificial Intelligence for Academic Performance Prediction. An Experimental Study on the Impact of Accuracy and Simplicity of Decision Trees on Causability and Fairness Perceptions

The rising adoption of learning analytics and academic performance prediction technologies in higher education highlights the urgent need for transparency and explainability. This demand, rooted in ethical concerns and fairness considerations, converges with Explainable Artificial Intelligence (XAI) principles. Despite the recognized importance of transparency and fairness in learning analytics, empirical studies examining student fairness perceptions, particularly within academic performance prediction, remain limited. We conducted a pre-registered factorial survey experiment involving 1,047 German students to investigate how decision tree features (simplicity and accuracy) influence perceived distributive and informational fairness, mediated by causability (i.e., the self-assessed understandability of a machine learning model’s cause-effect linkages). Additionally, we examined the moderating role of institutional trust in these relationships. Our results indicate that decision tree simplicity positively affects fairness perceptions, mediated by causability. In contrast, prediction accuracy neither directly nor indirectly influences these perceptions. Even if the hypothesized effects of interest are either minor or non-existent, results show that the medium positive effect of causability on the distributive fairness assessment depends on institutional trust. These findings substantially impact the crafting of transparent machine learning models in educational settings. We discuss important implications for fairness and transparency in implementing academic performance prediction systems.


INTRODUCTION
Higher education institutions are exploring academic performance prediction (APP) using machine learning (ML) in learning analytics (LA).Ensuring transparency and explainability is crucial, mandated by EU regulations and ethical guidelines, as well as student demand [43].This urgency is rooted in social fairness issues, where Explainable Artificial Intelligence (XAI) is relevant [1,34,36,46].Though transparency is emphasized in educational policies [14], "the number of empirical user studies examining ethical considerations, such as transparency in AI, is relatively low and often focused on LA in general rather than specific LA systems" [19].Our research thus focuses on students' perceptions of APP fairness.Using a factorial survey and a pre-registered experiment with 1,047 German students, we evaluated the impact of decision tree features-model simplicity and accuracy-on perceived fairness.We also examined the mediating role of causability, i.e., the self-assessed understandability of the cause-effect relationships of a ML model [40,78] in these relationships and assessed if institutional trust moderates these effects.

ACADEMIC PERFORMANCE PREDICTION AS A FORM OF AI-BASED LEARNING ANALYTICS
Within the last few years, many LA applications have been developed and adopted worldwide to support the work of students, lecturers, and administrators alike [16,84].Some of those applications can be used to reach several goals at once.For instance, with the implementation of APP systems, higher education institutions (HEI) aim to improve student success and reach higher equality in retention rates [5,67].APP systems are being used to predict students' performance based on large amounts of data-mainly historical performance data, but in some cases also sociodemographic data [2,27]-with the help of ML [4].The actual prediction can vary from the prediction of students' grades [6,20] over the likelihood of a successful study completion to the prediction of a potential dropout [12,64].Moreover, when APP is used to give individualized feedback or to distribute support measures, APP can further help especially those students who have often been disadvantaged before, by providing them with a more tailored educational experience that ensures that each student's unique needs and potential are recognized and addressed [60,67].

The Issues of Discrimination and Fairness Perceptions
However, discrimination and fairness are critical concerns in applying LA and APP [9].Discrimination can arise from flawed algorithmic design, biased data, or actions based on AI predictions [49,63].These issues can perpetuate societal biases [29,58] and influence human decision-making [49].Efforts exist to enhance the fairness of LA systems [26,42,61].Perceived fairness issues are equally important [46].APP applications are sociotechnical systems; thus, stakeholder perceptions, particularly students', are crucial [38,51,55].

On the Value of Explainable AI for APP Fairness
To achieve fairness and trust in LA, recent research has turned towards XAI methods [91].According to Gunning et al., "the purpose of an explainable AI (XAI) system is to make its behavior more intelligible to humans by providing explanations.(...) The XAI system should be able to explain its capabilities and understandings; explain what it has done, what it is doing now, and what will happen next; and disclose the salient information that it is acting on" [35].
Adadi and Berrada identify four reasons for the necessity of XAI: justification, control, improvement of AI systems, and knowledge production [1].XAI is deemed critical for individuals to understand and verify decisions [10] and is considered a prerequisite for fairness [34].Empirical evidence suggests that explanations improve perceived fairness and trust in AI systems [8,25,78,79,82].Specifically, explanations have shown to affect informational fairness [76,89].However, the impact varies depending on the dimensions of fairness and types of explanations used [11,25,75,77,79].Despite the demand for transparency, particularly from students [68,80,86], there is still a gap in understanding user prerequisites, needs and, expectations [18,30,47].In the context of LA, decision trees are often used for XAI in APP [41].These decision trees are trained on historical data, such as past exam results, to predict future academic outcomes [36,50].By making explanatory factors transparent, biases can be identified more quickly, contributing to the development of fairer algorithms [54].However, the complexity and accuracy of decision trees can vary [50].

Simplicity and Fairness Perceptions
As ML models grow in complexity, there is a risk they may become too intricate for human comprehension.Balancing complexity and simplicity is essential for explainability, although no definitive standards exist [74].Cognitive limitations further complicate the issue; for instance, young adults can process only three to five stimuli at the same time [21].Empirical studies reveal a nuanced relationship between informational fairness and the amount of explanation provided [23,48,76].However, the theory of explanatory coherence suggests that simpler explanations are generally preferred [62,83].This preference for simplicity has been empirically supported and observed in various applications, including health symptom checks and mathematical fairness notions [69,81,87].Yurrita et al. 's qualitative study indicated that too much information or complexity could be counterproductive, especially for those with limited AI literacy [89].Hence, the simplicity of an APP decision tree is critical to ensure equitable understanding and to prevent potential discrimination, such as favoring students with prior computer science knowledge.Based on these considerations, we formulate the first hypothesis: Hypothesis 1 (H1).Simpler decision trees lead to higher perceived informational fairness.

Accuracy and Fairness Perceptions
Accuracy is crucial for adopting and trusting APP systems.Low accuracy hinders adoption and impacts the system's fairness [37,49].Literature on algorithmic aversion indicates that observed errors in AI systems reduce people's confidence in them [24,57].Accuracy strongly predicts intention to follow AI recommendations, even more so than clarity of origin [32].It also positively affects trust in AI [65,66,88].However, Conijn et al. found no effect of accuracy explanations on student motivation in an essay grading context [19].Given the significance of accuracy and the inherent trade-offs in system design-since principles like transparency, explainability, and accuracy cannot be simultaneously maximized within a single ML model or XAI approach [3]-we formulate the second hypothesis: Hypothesis 2 (H2).Decision trees with a higher accuracy lead to higher perceived distributive fairness.

Fairness and Causability
Simplicity in APP systems aims to improve informational fairness but does not guarantee understandability for all students.If explanations are accessible only to a subset of students, unfairness ensues [10].In this regard, the concept of causability, introduced by Holzinger et al., assesses the quality of explanations from the user's perspective [40].Unlike explainability, which focuses on the system's capability to elucidate its functions [35], causability is usercentric and measures a person's understanding of an explanation.Holzinger et al. operationalized this with the system causability scale [39].Shin's study on AI journalism supports causability's role as an antecedent to explainability and its influence on perceived fairness and trust [78].Based on these insights, we formulate the third and fourth hypotheses: Hypothesis 3 (H3).The relationship between the simplicity of decision trees and perceived informational fairness is mediated by causability, such that simpler decision trees lead to greater causability, which in turn leads to increased perceived informational fairness.
Hypothesis 4 (H4).The relationship between the accuracy of decision trees and perceived distributive fairness is mediated by causability, such that decision trees with a higher accuracy lead to greater causability, which in turn leads to increased perceived distributive fairness.

Fairness and Institutional Trust
HEIs, as the stewards of APP systems, carry the ethical responsibility to safeguard students from the consequences of unfair decisions [45].Failure in this regard risks eroding institutional trust, as seen in cases involving automated assessment discrimination [28] and unethical data use [45].However, trust in HEIs significantly influences students' willingness to disclose data for LA [53,80].In the sociotechnical landscape of APP, trust extends beyond the technology to include the organizations and individuals that deploy it [56,85].Trust also acts as a complexity-reducing mechanism in uncertain situations [52].Therefore, students' inherent trust in their HEIs could potentially mitigate the need to fully comprehend APP's intricacies, assuming the institution is perceived as ethical and trustworthy [80,90].Based on these considerations, we formulate the fifth and sixth hypotheses: Hypothesis 5 (H5).The indirect effect of the simplicity of decision trees on perceived informational fairness through causability is moderated by institutional trust.Specifically, simpler decision trees lead to greater causability, which in turn results in increased perceived informational fairness.However, the strength of this mediated relationship is contingent upon the level of institutional trust.The indirect effect is weaker at higher levels of institutional trust, as individuals with high institutional trust have a higher perception of informational fairness even when simplicity and accuracy are low, making the role of causability as a mediator less influential in these cases.
Hypothesis 6 (H6).The indirect effect of the accuracy of decision trees on perceived distributive fairness through causability is moderated by institutional trust.Specifically, simpler decision trees lead to greater causability, which in turn results in increased perceived distributive fairness.However, the strength of this mediated relationship is contingent upon the level of institutional trust.The indirect effect is weaker at higher levels of institutional trust, as individuals with high institutional trust have a higher perception of distributive fairness even when simplicity and accuracy are low, making the role of causability as a mediator less influential in these cases.
Our primary focus is on assessing the impact of accuracy on distributive fairness and simplicity on informational fairness, with attention to the roles of causability and institutional trust.Our conceptual moderated mediation model is illustrated in Figure 1.In this model, two paths are not hypothesized to have direct effects.Subsequently, these paths will be freely estimated in our structural regression analysis, allowing for data-driven insights as this study aims to elucidate how simplicity and accuracy in decision trees affect fairness perceptions.Consequently, we pose two research questions:

RQ1. To what extent do simple decision trees affect perceived distributive fairness? RQ2. To what extent do accurate decision trees affect perceived informational fairness?
All hypotheses and research questions were pre-registered via OSF.

METHOD
In a 2x2 between-subjects experimental design, we examine the influence of decision tree simplicity and accuracy on students' causability and subsequent fairness perceptions.We also evaluate the moderating role of institutional trust on the causability-fairness relationship through a moderated mediation model.Data analysis was carried out using the statistical program R version 4.3.1 (2023-06-16 ucrt) using structural equation modeling with the package lavaan [73].All estimated models utilize bootstrapping with 5000 bootstraps.Bootstrap intervals for parameter estimates were produced using the adjusted bootstrap percentile method with bias correction.
The sample size for our study was determined through an a priori power analysis conducted in R, aiming for a power level of .75 with an anticipated sample size of around 1000 participants.This increased to approximately .8 with 1100 participants.We targeted a small effect size difference of .15 for the moderation of a mediation effect at a conventional alpha error probability of .05.Despite the risk of a 25% Type II error, we deemed this power level acceptable for our specific research context, balancing data collection feasibility with the robustness of our results.As the questionnaire needed to be accessed via a non-mobile device to display the stimulus correctly, the maximum number of respondents available via the panel provider was approximately 1100.Participants were recruited from the online panel Talk Online Data Collection AG, were 18 or older, and enrolled in higher education.The Ethical Review Board of the Faculty of Philosophy of the Heinrich Heine University Düsseldorf, Germany, approved the study.
All in all, n = 1047 students completed the survey.The average age of students was 25.46 (SD = 5.6).Altogether, 572 (54.7%) students identified as women, 462 (44.2%) as men, and 11 (1.1%) did not identify strictly as male or female, indicating 'diverse'.Of all students, 393 (37.5%) indicated studying a STEM subject and 606 (57.9%) indicated studying a non-STEM subject with 48 (4.6%) students indicating to study something else which could not be assigned to either STEM or non-STEM subjects.Regarding the question of which degree students are currently pursuing in their core subject, 636 (60.7%) students report pursuing a bachelor's degree, 246 (23.5%) report pursuing a master's degree, 129 (12.3%) report pursuing state examination, and 36 (3.4%) pursue a doctorate.

Procedure and Survey Design
First, respondents were briefed on the study's objectives, questionnaire duration, and data protection measures.After giving informed consent, they confirmed current enrollment in a HEI.Sociodemographic data and institutional trust levels were then collected.A brief overview of AI and Academic Performance Prediction APP was provided before randomly presenting one of four decision tree stimuli, varying in complexity and accuracy.A treatment check and questions about the decision tree's causability followed.Participants then evaluated the tree's distributive and informational fairness before answering additional academic-related questions.Finally, they were debriefed, redirected to the panel provider, and compensated.The average time to complete the questionnaire was 6.06 minutes (SD = 2.55).

Independent Variable (IV). Simplicity
As stimuli, students viewed one of four decision trees detailing factors affecting APP.The trees varied in simplicity, serving as an independent variable.The simpler tree had two decision levels leading to an outcome prediction, while the more complex one had up to five decision rules.These factors were based on a real APP system in development [27] and excluded potentially discriminatory data.Tailored to computer science and social sciences students, the factors were abstracted for cross-field comparability, allowing identification for respondents across subjects.
IV. Accuracy Regarding accuracy, the second independent variable, rates differed between a higher accuracy of 95% and a lower accuracy of 65%.This information was displayed conspicuously below the decision tree.These rates were chosen to capture performance variance.In the 2x2 between-subject design, students viewed a decision tree that was either simple or complex and had an accuracy of either 65% or 95%.For visual reference, see Figures 2 and 3.
As a dependent variable (DV), we focused on students' perceptions of fairness.Thus, information fairness was measured with several items in the first step.To achieve good factorial validity (Cronbach's  = 0.80; AVE = 0.57), we decided to choose the following three items for the latent variable of our structural equation model: "The reasons for the prediction are understandable."; "The explanation of the AI-based performance prediction procedure is comprehensive.";"The explanation of the prediction is coherent."The first item is self-developed, while the other two were adapted from Schoeffer et al. [77].All variables used for the study were measured on a five-point Likert scale ranging from 1 = "strongly disagree" to 5 = "strongly agree".Respondents also had the option of expressing no preference ("don't know"; except in the case of institutional trust). 1V.Distributive Fairness.
Students' perceived distributive justice was measured using three commonly used items developed by Colquitt and Rodell [17] and adapted to the APP context.Students rated whether they agreed with the following statements: "The prediction of performance for students by AI is fair."; "Everyone gets what he/she deserves."; "No one is unduly disadvantaged by student performance prediction by AI. " All three items also show good factorial validity (Cronbach's  = 0.83; AVE = 0.62).
With the two dependent variables, informational fairness and distributive fairness, demonstrating good convergent validity, the question arises as to what extent they are distinct constructs, addressing the issue of discriminant validity.A high correlation is observable in a model where both constructs are estimated as individual latent factors, r = 0.68.However, discriminant validity assessment does not only focus on the observed correlation between the constructs.By applying the Fornell-Larcker criterion [31], which stipulates that the squared correlation between the constructs must be less than the individual AVE of each factor, results suggest that there is discriminant validity,  2 = 0.46.

Mediator. Causability
To assess students' understanding of the APP presented, namely causability, we used the System Causability Scale from Holzinger et al. [39].Again, however, we had to exclude items to improve factorial validity.Respondents were asked to indicate how much they agreed with the following statements regarding the explanations in APP's decision tree: "I understand the explanations in the context of my studies."; "The explanations help me understand the criteria of the prediction."; "I am able to understand the explanations with my prior knowledge."Previously, it was explained that the term "explanations" refers to the decision process shown earlier, which shows at the end whether a dropout is predicted by the AI system or not.The items show good factorial validity (Cronbach's  = 0.78; AVE = 0.54).
Moderator.Institutional Trust.Finally, institutional trust was measured regarding the HEI at which students are enrolled.We used and slightly adapted four items developed by Gosh et al. [33] and previously validated by Li et al. [53].The students were asked to rate the extent to which they agreed with the following sentences: "Since I cannot personally supervise all of my university's activities, I rely on the university staff to do their jobs properly.";"I believe that my university is a credible organization.";"I feel that I can rely on my university.".However, to improve factorial validity, we excluded the reverse coded item ("In general, I do not have confidence in my university.").After doing so, the items show good factorial validity (Cronbach's  = 0.75; AVE = 0.53).

Treatment Check
Two treatment check items assessed participants' perceptions of the manipulated conditions.Participants responded on a five-point scale, ranging from 1 (strongly disagree) to 5 (strongly agree).The first item examined the perception of decision tree simplicity: If you think about the decision tree just shown: To what extent do you agree or disagree with the following statements?The decision tree of the performance prediction is straightforward.The results indicated a significant difference among the four conditions (F(3, 577.9) = 42.8,p < 0.001).Using a Games-Howell post hoc test, the conditions with a less simple decision tree and high accuracy (M = 3.40; SD = 1.11) and lower accuracy, respectively (M = 3.45; SD = 1.16), were found to differ significantly from the conditions with a simpler decision tree and high accuracy (M = 4.17; SD = 0.93) and lower accuracy, respectively (M = 4.14; SD = 0.96), confirming that respondents recognized the extent to which the displayed simplicity of the decision tree varied.The second item assessed the perception of decision tree accuracy: The performance prediction decision tree shows high accuracy.The results indicated a significant difference among the four conditions (F(3, 1043) = 20.34,p < 0.001).Using a Games-Howell post hoc test, the conditions with high accuracy and a less simple decision tree (M = 3.31; SD = 1.05) and simpler decision tree, respectively (M = 3.36; SD = 1.06), were found to differ significantly from the conditions with low accuracy and a less simple decision tree and (M = 2.77; SD = 1.08) and simpler decision tree, respectively (M = 2.89; SD = 1.08), confirming that respondents recognized the extent to which the displayed accuracy of the decision tree varied.

Main Effects Model
We first address a main effects model to test H1 and H2, reporting the direct effects of the exogenous variables simplicity and accuracy that were manipulated in the stimulus as IVs on the DVs informational and distributive fairness.For the sake of completeness, in this analysis, we estimated a model that also includes the interaction of simplicity and accuracy.The structural regression model shows good fit ( 2 (20) = 68.89,p = < 0.001; RMSEA = 0.05, 90% CI [0.04, 0.06]; TLI = 0.97).
Regarding the simplicity of the decision tree, results suggest that there is a small positive and significant effect of the simplicity of the decision tree on perceptions of informational fairness; the simpler the decision tree, the greater the perceived informational fairness, B = 0.32, SE = 0.10, 95% CI (0.14, 0.51), p = < 0.001,  = 0.16.Accordingly, H1 is accepted.Additionally, there is a small positive and significant effect of simplicity on perceived distributive fairness; the simpler the decision tree, the greater the perceived distributive fairness, B = 0.31, SE = 0.12, 95% CI (0.07, 0.52), p = 0.008,  = 0.12.

Mediation Model
Second, to test H3 and H4, we estimated a mediation model that integrates the mediator, causability, to understand how the IVs affect the DVs through an indirect pathway.Specifically, the model not only tests whether simplicity and accuracy directly affect both informational and distributive fairness but also influence them indirectly via causability.The structural regression model shows good fit ( 2 (36) = 99.30,p = < 0.001; RMSEA = 0.04, 90% CI [0.03, 0.05]; TLI = 0.98).
Regarding the effects of the IVs on the mediator, causability (i.e., the first path of the indirect effect), we first examine the impacts of simplicity and accuracy.For the simplicity of the decision tree, there was positive effect on causability; the simpler the decision tree, the greater the causability.The effect is small and significant, B = 0.16, SE = 0.07, 95% CI (0.03, 0.29), p = 0.016,  = 0.08.For the accuracy of the decision tree, there was no significant effect on causability, B = 0.04, SE = 0.07, 95% CI (-0.10, 0.17), p = 0.552,  = 0.02.
Second, we assess the effects of the mediator, causability, on the DVs informational fairness and distributive fairness (i.e., the second path of the indirect effect).First, there was positive effect of causability on informational fairness; the higher the causability, the greater the perceived informational fairness.The effect is strong and significant, B = 0.72, SE = 0.05, 95% CI (0.62, 0.82), p = < 0.001,  = 0.69.Second, there was positive effect of causability on distributive fairness; the higher the causability, the greater the perceived distributive fairness.The effect is strong and significant, too, B = 0.55, SE = 0.05, 95% CI (0.45, 0.66), p = < 0.001,  = 0.43.

Moderated Mediation Model
Lastly, to test H5 and H6, we estimated a moderated mediation model that integrates the moderator, institutional trust, to test whether the effect of causability on informational fairness and distributive fairness as part of the indirect effect of the IV on the DV via the mediator depends on the extent of students' institutional trust.We first report a model with parameter estimates that treat the moderator as a continous latent variable.Second, we compare the model across three groups to illustrate the changes in the effect at different levels of the moderator: a) a group of students whose institutional trust scores are at least one standard deviation (SD) below the mean, b) a group whose institutional trust is within one SD of the mean, and c) a group whose institutional trust is one SD above the mean.
The first estimated model, which treats the moderator as a continuous latent variable, suggests a poor fit ( 2 (213) = 1851.31,p = < 0.001; RMSEA = 0.09, 90% CI [0.08, 0.09]; TLI = 0.81).This poor fit arises from the introduction of the interaction terms, which multiply both the amount and magnitude of covariances between variables that are assumed to be independent.When combined with the sample size, this increase in covariance elevates the chi-square value, a measure of goodness of fit, possibly leading to rejection of the model under conventional thresholds.However, the model does reach a plausible solution, and an examination of the parameter estimates indicates no major deviation compared to the previously estimated models.
Figure 4 below shows the parameter estimate for the effect of causability on informational fairness for students at least one SD below the mean of institutional trust, within one SD of the mean, and at least one SD above the mean.As there is no moderation effect, we reject H5.
The parameter estimate for second interaction effect suggests a small negative and significant effect of institutional trust on the relationship between the causability and distributive fairness, B = -0.11,SE = 0.06, 95% CI (-0.21, 0.01), p = 0.054,  = -0.07.
Figure 5 below shows the parameter estimate for the effect of causability on distributive fairness for students at least one SD below the mean of institutional trust, within one SD of the mean, and at least one SD above the mean.While our confidence intervals for the individual parameter estimates in the figure do overlap, suggesting uncertainty in the distinctions between some groups, our more direct tests of differences between specific groups show significant contrasts.For instance, the difference of the regression parameter 'Distributive Fairness ~Causability' between 'More than 1 SD below Mean' and 'More than 1 SD above Mean' is statistically significant, B = 0.42, SE = 0.21, 95% CI (0.04, 0.86), p = 0.046, Δ = 0.29.This suggests that even though the confidence bands for these groups might overlap when calculated and visualized individually, the statistical evidence points towards a difference in their actual parameter estimates.However, as there was overall no total effect of accuracy on distributive fairness and no indirect effect via the causability, the different indirect effects for the first group, B = -0.00,SE = 0.14, 95% CI (-0.26, 0.28), p = 0.981,  = -0.00, the second group, B = 0.01, SE = 0.05, 95% CI (-0.08, 0.09), p = 0.831,  = 0.00, and the third group, B = 0.09, SE = 0.07, 95% CI (-0.04, 0.26), p = 0.236,  = 0.03, are too small to reach significance and suggest practically no effect for fairness perceptions.Accordingly, H6 is rejected.

DISCUSSION
Concerning H1 and H3, simplicity in decision trees for APP positively influenced perceptions of informational fairness mediated by causability.This is consistent with explanatory coherence theory [83] suggesting more simple explanations are preferred, and it is thus crucial to balance providing sufficient information and avoiding overwhelming students with excessive details.Confirming H3, we find evidence for the assumption that more information does not necessarily lead to a better understanding and higher fairness perceptions by default [7,23], highlighting the importance of students' self-assessed understandability of the decision trees that visualized the ML model's outcome [40].Moreover, H5, positing institutional trust as a moderator, was not supported, highlighting that trust cannot replace individual comprehension to ensure informational fairness [80].Regarding RQ1, simplicity had a non-hypothesized positive effect on informational and distributive fairness mediated by causability.This supports existing literature arguing that overall simpler explanations are favored [69,81,83,87].
H2 and H4, which posited that accuracy would impact distributive fairness and be mediated by causability, were not supported.This raises questions about students' awareness of the risks associated with low-accuracy AI.The absence of an effect from a 30-percentage point accuracy manipulation is notable but may also be linked to external validity, as respondents faced no real-world consequences from model errors.Like informational fairness, causability was a strong determinant in distributive fairness, emphasizing its key role in perceptions of fair AI decisions.In the absence of both the overall effects of accuracy on distributive fairness and the indirect effects via causability, H6 had to be rejected.However, institutional trust did affect the relationship between causability and distributive fairness.This suggests that higher institutional trust reduces the importance of the decision tree's understandability on perceptions of distributive fairness.Accordingly, institutional trust may still contribute to APP being judged fairly even if the AI system is not entirely understood, as there can be trust that the university will act ethically and respect students' interests [45,80].Lastly, as questioned in RQ2, accuracy also shows no influence on informational fairness.This finding remains constant when adding causability as a mediator in the model, indicating no significant effect.While the concrete level of accuracy of the performance of the APP model does not help to increase informational fairness, this does not mean that the knowledge about the APP's accuracy per se is not important in terms of system transparency.However, since Conijn et al. found that accuracy has no effect on student motivation or confidence in an essay grading system [19], one might assume that accuracy is more critical to the question of whether APP is good enough to be used at all than to the question of whether accuracy increases informational fairness perceptions.
Overall, the results underscore the limited impact of objective attributes like simplicity and accuracy on fairness perceptions, emphasizing the role of subjective factors like causability.Future research should focus on these subjective perceptions to better understand fairness in APP.

IMPLICATIONS AND CONCLUSION 6.1 Practical Implications
Our study finds that the design features of the decision tree had a limited impact on its perceived comprehensibility (i.e., causability) and fairness.However, design decisions remain critical for fair and effective communication in XAI.Relevant, high-quality explanations are essential for understanding the APP process and its  outcomes.Yet, these should be presented without overwhelming complexity.Our findings suggest that decision trees with two to five levels do not adversely affect self-assessed understanding or perceived fairness.Overall, respondents indicated that they understood the cause-effect relationships in the white-box model.It would be interesting to explore whether the complexity of other approaches, such as random forest classifiers, could positively or negatively impact causability.However, it is vital to tailor the complexity of explanations to the audience's cognitive abilities and prior knowledge, especially in diverse educational contexts.Failure to do so risks exacerbating inequalities by favoring those with a greater prior understanding of AI systems.It would thus be premature to dismiss the demands for accuracy regarding these models, as high accuracy is equated with the reliability and trustworthiness of AI systems.Nonetheless, it is essential to acknowledge that wellintentioned policy changes to improve system transparency and fairness may not resonate across all audience segments by default.In sum, the key is providing explanations and ensuring they are perceived as comprehensible to those they affect.

Research Implications
Our study points to the nuanced roles of simplicity and accuracy in shaping fairness perceptions in APP, urging further investigation into other XAI attributes [71].For example, how counterintuitive factors in decision trees, like high grades predicting poor performance, influence student perceptions remains an open question.The study also calls for understanding how varying student attributes affect the comprehensibility of explanations, raising the issue of whether a one-size-fits-all explanation is adequate.Further research is needed to ascertain the optimal level of explanation that avoids information overload in increasingly complex models.Moreover, the role of universities in APP implementation requires further examination.While institutional trust is vital, it cannot replace the need for individual comprehension of LA explanations.Our findings suggest that decision tree simplicity positively affects fairness perceptions, mediated by causability, whereas prediction accuracy has a less pronounced impact.Our study thus offers critical insights for stakeholders in HEIs, highlighting the importance of balancing explainability and comprehensibility to foster ethical and equitable practices in academic settings.
In conclusion, our study underscores the importance of explainability and comprehensibility in implementing LA and APP technologies in HEIs.Decision tree simplicity emerges as a factor positively influencing fairness perceptions, mediated by causability, while prediction accuracy appears to play a less significant role in shaping student perceptions of fairness.Importantly, our findings highlight the need to explore further institutional trust's multifaceted influence on fairness perceptions in academic contexts.As HEIs strive for fair and transparent APP systems, our research offers valuable insights for policymakers, administrators, and students, fostering ethical and equitable practices in higher education.

ACKNOWLEDGMENTS Ethics Statement
In conducting this research, we adhered to the highest standards of ethical integrity and responsibility.We ensured that all participants were fully informed about the nature and purpose of the research and provided their informed consent.Participant confidentiality and data privacy were rigorously maintained throughout the study.Ethical guidelines, including those pertaining to nondiscrimination, fairness, and respect for individuals, were strictly followed.The research methods were designed to minimize potential harm or discomfort to participants.Any conflicts of interest were disclosed and managed appropriately.This study received approval from the Ethical Review Board of the Faculty of Philosophy of the Heinrich Heine University Düsseldorf, Germany, and was conducted in accordance with international standards for ethical research, such as the Helsinki Declaration.

Figure 2 :Figure 3 :
Figure 2: Simple Decision Tree for Academic Performance Prediction with High Accuracy

Figure 4 :
Figure 4: The Effect of Causability on Informational Fairness Depending on Institutional Trust

Figure 5 :
Figure 5: The Effect of Causability on Distributive Fairness Depending on Institutional Trust