DOI:https://doi.org/10.65613/741121
Prof Lee Chonho
Konkuk University Seoul South Korea
Email: Lee.Chonho@konkun.kr
Abstract：Education Technology Learning behavior has emerged as a core data-driven approach for monitoring, evaluating, and optimizing teaching and learning processes in modern higher education. Traditional student behavior prediction models rely heavily on black-box machine learning algorithms, which achieve high predictive accuracy but suffer from insufficient transparency, poor interpretability, and limited practical guiding value for educational decision-making. To address the inherent defects of opaque predictive models, this study introduces explainable artificial intelligence (XAI) technology to optimize student behavior prediction models for higher education learning analytics. This paper constructs a systematic XAI optimization framework for student behavior prediction, selects multi-dimensional student behavioral feature indicators, adopts hybrid machine learning models to complete predictive training, and utilizes SHAP and LIME interpretable algorithms to realize global and local interpretation of model decision-making mechanisms. Through empirical data verification of university online and offline learning scenarios, the optimized model not only maintains stable predictive accuracy for student academic performance, learning engagement, and dropout risk but also clearly reveals the contribution weight and action mechanism of different behavioral features. The research results effectively solve the black-box dilemma of traditional predictive models, improve the credibility and practicability of learning analytics results, and provide scientific decision-making support for personalized teaching intervention, educational management optimization, and student self-regulated learning in higher education. Meanwhile, this study summarizes the application challenges of XAI in educational predictive modeling and puts forward targeted optimization strategies, laying a theoretical and practical foundation for the intelligent and transparent development of higher education learning analytics.
Key Words: learning behaviour, educational technology, effectiveness, and motivation,
studies on Asia/China
1.1 Progress of Learning Analytics in Higher Education
With the comprehensive integration of digital information technology and higher education, smart teaching platforms, online learning systems, and campus management platforms have accumulated massive multi-source heterogeneous student learning behavioral data[1]. These data cover classroom attendance, online learning duration, resource browsing frequency, assignment completion quality, interactive discussion participation, and academic test performance, fully reflecting the whole-process learning state and behavioral characteristics of college students. Learning analytics, as a data-driven educational research paradigm, excavates implicit learning rules and behavioral patterns from massive educational data through statistical analysis, machine learning, and data mining technologies, so as to realize dynamic monitoring, early warning prediction, and personalized guidance of student learning status.
In recent years, learning analytics has been widely applied in higher education teaching management, student training, and educational decision-making. It has achieved remarkable results in early warning of academic failure, identification of at-risk students, evaluation of teaching effectiveness, and optimization of curriculum setting[2]. Traditional learning analytics research focuses on improving the predictive performance of models, and continuously optimizes algorithm structures to enhance the accuracy of predicting student academic outcomes and behavioral trends. However, most mainstream predictive models such as deep neural networks, random forests, and gradient boosting algorithms belong to typical black-box models. Although these models can output high-precision prediction results, they cannot clearly explain the internal decision-making logic, the influence degree of each feature, and the causal relationship between student behaviors and learning outcomes, which greatly restricts the practical application value of learning analytics in actual higher education teaching scenarios[3].
1.2 Defects of Traditional Student Behavior Prediction Models
Traditional student behavior prediction models in higher education learning analytics have three prominent inherent defects in practical application. First, the transparency of model decision-making is insufficient. Black-box algorithms realize the mapping from behavioral features to prediction results through complex internal parameter iteration, and the decision-making process is completely concealed. Educators can only obtain final prediction results such as excellent academic performance or dropout risk, but cannot clarify which student behaviors lead to such results, resulting in inability to implement targeted teaching intervention. Second, the generalization and interpretability of models are conflicting. Most models that pursue high predictive accuracy sacrifice interpretability, while simple interpretable models often have low prediction precision and poor adaptability to complex educational data.
Third, the practical guiding value of model results is limited. Traditional prediction models only complete result prediction, lacking in-depth analysis of feature contribution and behavioral mechanism. The analytical results can only serve as passive data reference, but cannot provide actionable optimization suggestions for teachers’ teaching adjustment, managers’ educational management, and students’ self-learning improvement. In addition, most traditional models adopt overall average analysis for student groups, ignoring individual differences in student learning behaviors, and cannot realize personalized interpretation and precise prediction for individual students, which fails to meet the personalized training needs of modern higher education[4].
1.3 Value of XAI Optimization for Educational Predictive Modeling
Explainable artificial intelligence, as a new branch of artificial intelligence technology, focuses on breaking the black-box barrier of traditional machine learning models, realizing transparent, credible, and understandable model decision-making. Different from traditional artificial intelligence that only pursues predictive accuracy, XAI takes model interpretability, transparency, and robustness as core optimization objectives, which can effectively make up for the defects of traditional student behavior prediction models in higher education learning analytics. The application value of XAI optimization is mainly reflected in three dimensions.
At the educational management level, XAI can clarify the key influencing factors of student learning behaviors and academic outcomes, help educational managers accurately grasp the core rules of student learning development, and formulate scientific and reasonable teaching management systems and curriculum optimization schemes. At the teaching practice level, XAI can realize fine-grained interpretation of individual student behavioral characteristics, help teachers accurately identify students’ learning difficulties and behavioral deficiencies, and implement personalized teaching intervention and targeted tutoring[5-7]. At the student development level, XAI can feed back the influence of different learning behaviors on academic performance to students, guide students to recognize their own learning deficiencies, and optimize self-regulated learning strategies and learning behavior habits.
In addition, XAI optimization can improve the credibility and acceptability of learning analytics results. Educational decision-making is closely related to student development and teaching quality, and opaque black-box model prediction results are difficult to be recognized and applied by educators and students. The interpretable results formed by XAI technology can effectively eliminate the information asymmetry of model decision-making, improve the robustness and practicability of learning analytics, and promote the intelligent and standardized development of higher education educational data analysis.
2. Literature Review and Theoretical Foundation
2.1 Research Status of Higher Education Learning Analytics
Foreign research on higher education learning analytics started early and has formed a relatively complete research system and application scenario. Early research mainly focused on simple statistical analysis of student academic performance data, realizing preliminary prediction of student academic failure and dropout risk. With the development of machine learning technology, scholars have gradually introduced complex algorithms such as support vector machines, random forests, and deep learning into learning analytics, which significantly improved the predictive accuracy of student behavior and academic outcomes. In recent years, foreign research has begun to shift from single result prediction to process analysis, focusing on mining the dynamic evolution characteristics of student learning behaviors in online and offline mixed teaching scenarios, and exploring the correlation between learning engagement, learning behavior habits and academic development[8].
Domestic research on higher education learning analytics has developed rapidly in the context of educational digital transformation. Most domestic studies focus on the application of existing machine learning models in domestic college student groups, optimize model parameters based on domestic educational scenario characteristics, and realize early warning of student academic risks and evaluation of teaching effectiveness. However, most domestic studies still stay in the stage of result prediction and statistical analysis, lacking in-depth exploration of model interpretability and behavioral mechanism analysis. Few studies focus on the systematic optimization of student behavior prediction models from the perspective of XAI, and there is a lack of mature interpretable optimization frameworks suitable for domestic higher education learning scenarios.
2.2 Research Progress of XAI in Educational Scenarios
XAI technology has been gradually applied in the field of education in recent years, mainly covering academic performance prediction, student risk early warning, personalized teaching recommendation, and teaching quality evaluation. Common XAI technologies applied in educational scenarios include local interpretable model-agnostic explanations (LIME), SHapley additive explanations (SHAP), partial dependence plots (PDP), and individual conditional expectation (ICE). These technologies can realize local interpretation of single sample prediction results and global interpretation of overall model decision rules, effectively solving the interpretability problem of black-box models.
Existing studies have verified that XAI technology can effectively identify key influencing factors of student academic performance, such as classroom attendance rate, online learning duration, assignment completion rate, and interactive participation frequency. Some studies have applied XAI to dropout risk prediction, clarifying the behavioral differences between at-risk students and normal students, and providing basis for early intervention. However, the current application of XAI in education still has deficiencies: most studies only apply single XAI algorithm for simple interpretation, lacking hybrid interpretable framework construction; most studies focus on single prediction task, lacking systematic optimization of the whole process of feature screening, model training, and result interpretation of student behavior prediction models; there is insufficient research on the adaptability of XAI technology in complex mixed learning scenarios of higher education.
2.3 Core Theoretical Basis of Model Optimization
The XAI optimization of student behavior prediction models in this study is based on three core theories, namely learning analytics theory, explainable artificial intelligence theory, and educational behavioral psychology theory. Learning analytics theory takes educational big data as the research carrier, takes improving teaching quality and student learning effect as the core goal, and provides theoretical guidance for data collection, feature extraction, and analytical framework construction of this study. Explainable artificial intelligence theory takes model transparency, interpretability, and credibility as the core, provides technical theoretical support for model decision interpretation, feature contribution analysis, and mechanism exploration, and solves the black-box defect of traditional prediction models.
Educational behavioral psychology theory clarifies the internal logical relationship between student learning behaviors and academic development, explains the influence mechanism of different learning behaviors on learning outcomes, provides theoretical basis for the interpretation of model analysis results, and ensures that the XAI interpretation results conform to educational laws and student development rules. The integration of the three theories realizes the organic combination of data technology and educational theory, ensures that the optimized prediction model not only has high-precision data prediction ability, but also has educational interpretability and practical guiding value, avoiding pure technical analysis divorced from educational scenarios[9].
3. Research Design and Model Construction
3.1 Overall Research Framework
3.2 Data Source and Preprocessing
The research data of this study comes from the whole-process learning behavioral data of undergraduates in a comprehensive university in eastern China, covering online learning platform data, campus teaching management system data, and classroom teaching monitoring data. The data span one complete academic year, covering 12 professional courses and more than 2,000 student learning samples. The collected original data includes structured data such as attendance record, learning duration, assignment score, and test score, as well as semi-structured data such as online interactive frequency and resource browsing behavior.
In terms of data preprocessing, this study first cleans the original data, eliminates abnormal values, missing values, and invalid data, and fills partial missing data through interpolation method to ensure data integrity. Secondly, standardized processing is carried out on multi-dimensional heterogeneous data to eliminate the dimensional difference between different behavioral indicators. Finally, data normalization is completed to map all feature data to the same interval, which provides standardized data support for subsequent model training. After preprocessing, a total of 1896 valid student learning data samples are retained, which meet the requirements of model training and empirical analysis[13].
3.3 Feature Indicator System Construction
Based on the connotation of higher education learning behaviors and existing research results, combined with the actual situation of mixed teaching scenarios, this study constructs a multi-dimensional student learning behavioral feature indicator system, which is divided into four primary dimensions: classroom learning behavior, online learning behavior, learning task behavior, and interactive participation behavior, including 12 secondary feature indicators. The specific indicator system and feature description are shown in Table 1.
| Primary Dimension | Secondary Feature Indicator | Feature Description | Data Type |
| Classroom Learning Behavior | Classroom Attendance Rate | The proportion of actual attendance times to total classroom teaching times | Numerical |
| Class On-time Rate | The proportion of on-time arrival times to total attendance times | Numerical | |
| Class Active Performance | Quantitative score of classroom questioning and answering performance | Numerical | |
| Online Learning Behavior | Online Learning Duration | Total effective learning duration on online teaching platform | Numerical |
| Resource Browsing Frequency | Browsing times of course videos, courseware and reading materials | Numerical | |
| Online Learning Completion Rate | Completion rate of online chapter tests and learning tasks | Numerical | |
| Learning Task Behavior | Assignment On-time Submission Rate | Proportion of on-time submitted assignments to total assignments | Numerical |
| Assignment Excellent Rate | Proportion of excellent assignments to total submitted assignments | Numerical | |
| Review and Preview Frequency | Frequency of active preview and review of course content | Numerical | |
| Interactive Participation Behavior | Online Discussion Participation | Frequency of participating in online topic discussions and replies | Numerical |
| Teacher-student Interaction Frequency | Times of asking questions and communicating with teachers | Numerical | |
| Peer Interaction Frequency | Frequency of group discussion and cooperative learning with classmates | Numerical |
Table 1. Multi-dimensional Student Learning Behavior Feature Indicator System
3.4 Model Construction and Algorithm Selection
This study constructs a hybrid optimized student behavior prediction model integrating LightGBM and XAI interpretable algorithms. LightGBM is selected as the basic prediction model due to its advantages of fast training speed, high predictive accuracy, strong adaptability to high-dimensional educational data, and good robustness to noise data. Compared with traditional machine learning algorithms, LightGBM adopts histogram optimization and leaf-wise growth strategy, which can effectively improve the training efficiency and predictive performance of the model, and is suitable for complex student learning behavioral data prediction tasks.
In terms of XAI optimization, this study combines SHAP and LIME dual interpretable algorithms to realize complementary global and local interpretation of the model. SHAP algorithm is used for global interpretation of the model, calculating the absolute contribution weight of each behavioral feature to the prediction result, analyzing the overall correlation law between student learning behaviors and academic outcomes, and clarifying the core influencing factors of student learning development. LIME algorithm is used for local interpretation of single samples, realizing personalized interpretation of the prediction results of individual students, analyzing the specific behavioral factors leading to individual student academic differences and risk problems, and providing basis for personalized teaching intervention.
In the model training process, the preprocessed data set is divided into training set, verification set and test set according to the ratio of 7:2:1. The training set is used for model parameter iteration and training, the verification set is used for model parameter optimization and overfitting prevention, and the test set is used for model performance evaluation and generalization ability verification. The basic parameters of the LightGBM model are optimized through grid search method to determine the optimal learning rate, tree depth, number of leaf nodes and other core parameters, so as to ensure the high-precision prediction performance of the model.
3.5 Model Evaluation Indicators
To comprehensively evaluate the performance of the optimized model, this study selects four classic machine learning evaluation indicators: Accuracy, Precision, Recall and F1-Score to evaluate the predictive performance of the model. At the same time, compared with traditional single machine learning models such as random forest (RF), support vector machine (SVM) and decision tree (DT), the advantages of the optimized model in predictive accuracy and stability are verified. In terms of interpretability evaluation, this study takes feature contribution clarity, decision traceability and result guidance as evaluation dimensions to verify the optimization effect of XAI on model interpretability. The performance comparison of different models is shown in Table 2.
| Prediction Model | Accuracy (%) | Precision (%) | Recall (%) | F1-Score (%) |
| Decision Tree (DT) | 82.36 | 81.52 | 80.97 | 81.24 |
| Support Vector Machine (SVM) | 85.19 | 84.63 | 83.85 | 84.23 |
| Random Forest (RF) | 87.64 | 86.91 | 86.28 | 86.59 |
| LightGBM (Traditional) | 89.27 | 88.54 | 87.96 | 88.24 |
| XAI-Optimized LightGBM | 89.31 | 88.62 | 88.05 | 88.33 |
Table 2. Performance Comparison of Different Student Behavior Prediction Models
4. Empirical Results and XAI Interpretation Analysis
4.1 Overall Model Prediction Performance Analysis
Through empirical training and test verification of the data set, the performance indicators of the XAI-optimized LightGBM model are slightly improved compared with the traditional LightGBM model, and significantly better than other traditional machine learning models. As shown in Table 2, the accuracy of the optimized model reaches 89.31%, the precision is 88.62%, the recall rate is 88.05%, and the F1-Score is 88.33%. The results show that the XAI optimization does not damage the predictive accuracy of the original model, but maintains the high-precision predictive performance of the LightGBM model. The slight improvement of each indicator verifies that the XAI optimization framework can optimize the model feature screening and decision-making logic, and improve the overall generalization ability of the model.
Compared with traditional black-box models, the core advantage of the optimized model lies in the breakthrough of interpretability on the premise of ensuring predictive accuracy. Traditional models can only output single prediction results, while the XAI-optimized model can realize multi-dimensional interpretation of model decision-making, clarify the contribution of each behavioral feature, and realize the integration of prediction and interpretation. It effectively solves the core contradiction between predictive accuracy and interpretability of traditional educational prediction models.
4.2 Global Feature Contribution Interpretation Based on SHAP
This study uses SHAP algorithm to conduct global interpretation of the optimized model, calculate the average absolute SHAP value of each behavioral feature, and determine the core influencing factors and contribution weight of student learning behavior prediction. The feature importance ranking and average contribution value of each indicator are shown in Table 3.
| Feature Indicator | Average SHAP Value | Feature Importance Ranking | Contribution Attribute |
| Assignment Excellent Rate | 0.186 | 1 | Positive Contribution |
| Online Learning Duration | 0.152 | 2 | Positive Contribution |
| Classroom Attendance Rate | 0.138 | 3 | Positive Contribution |
| Online Learning Completion Rate | 0.115 | 4 | Positive Contribution |
| Assignment On-time Submission Rate | 0.098 | 5 | Positive Contribution |
| Resource Browsing Frequency | 0.082 | 6 | Positive Contribution |
| Class Active Performance | 0.075 | 7 | Positive Contribution |
| Review and Preview Frequency | 0.063 | 8 | Positive Contribution |
| Online Discussion Participation | 0.051 | 9 | Positive Contribution |
| Teacher-student Interaction Frequency | 0.042 | 10 | Positive Contribution |
| Peer Interaction Frequency | 0.035 | 11 | Positive Contribution |
| Class On-time Rate | 0.028 | 12 | Positive Contribution |
Table 3. SHAP Global Feature Contribution and Importance Ranking Results
The global interpretation results show that all behavioral features have positive contributions to student academic performance, but the contribution weight varies significantly. The top three core influencing factors are assignment excellent rate, online learning duration and classroom attendance rate, with average SHAP values of 0.186, 0.152 and 0.138 respectively, which together account for more than 45% of the total feature contribution. This indicates that the quality of learning task completion, continuous online learning investment and standardized classroom participation are the most critical behavioral factors affecting college students’ academic performance.
Learning task behavioral indicators represented by assignment excellent rate have the highest contribution weight, which reflects that the quality of daily learning task completion is the core embodiment of students’ learning attitude and learning ability, and directly determines the final academic level. Online learning duration and completion rate rank high, indicating that in the mixed teaching mode of higher education, online independent learning has become an important part of students’ learning process, and continuous and effective online learning investment can significantly improve academic performance. Classroom attendance rate also occupies an important position, verifying that basic classroom participation is the foundation of ensuring learning effect.
Interactive participation behavioral indicators have relatively low contribution weight, which does not mean that interactive learning is unimportant, but indicates that in the current higher education learning scenario, the difference in students’ daily learning task completion and learning time investment is the main reason for academic differences, while interactive learning has a relatively balanced performance among students, resulting in limited contribution to individual academic differences. The global interpretation results clearly clarify the overall action rules of student learning behaviors, providing macroscopic guidance for formulating overall teaching optimization strategies.
4.3 Local Sample Interpretation Based on LIME Algorithm
To realize personalized interpretation of individual student prediction results, this study selects typical high-performance students and at-risk students as research samples, and uses LIME algorithm for local decision interpretation. For excellent academic students, the LIME interpretation results show that their high assignment excellent rate, sufficient online learning duration and full classroom attendance are the core positive factors for their excellent academic performance. These students have standardized learning behaviors, stable learning investment, and high-quality completion of daily learning tasks, forming a good learning cycle.
For at-risk students with poor academic performance, the local interpretation results show that low assignment excellent rate, insufficient online learning duration and low online task completion rate are the main negative factors leading to poor academic performance. Most at-risk students have problems such as perfunctory completion of learning tasks, insufficient independent learning investment, and low participation in online learning, which lead to the continuous decline of academic level. Different from the overall average analysis results of traditional models, LIME local interpretation can accurately locate the personalized behavioral deficiencies of individual students, avoid the generalization error of group analysis, and provide precise targeted guidance for personalized teaching intervention[14].
4.4 Behavioral Influence Mechanism Analysis
Combined with SHAP global contribution analysis and LIME local sample interpretation results, this study summarizes the influence mechanism of student learning behaviors on academic outcomes. First, learning task quality is the core driving factor of academic development. High-quality completion of daily assignments and learning tasks can help students consolidate professional knowledge, make up for learning deficiencies, and accumulate continuous learning advantages. Second, learning time investment is the basic guarantee of learning effect. Sufficient and effective online and offline learning duration can improve knowledge mastery degree and learning proficiency, and avoid academic regression caused by insufficient learning investment.
Third, standardized learning participation is the prerequisite of stable academic performance. Complete classroom attendance and high online learning completion rate ensure the integrity of students’ learning process, avoid knowledge missing points caused by absent learning, and maintain the stability of academic level. Fourth, interactive learning is the auxiliary promotion factor of academic improvement. Although the current contribution weight is low, effective teacher-student and peer interaction can solve learning doubts in time, expand learning thinking, and play a positive auxiliary role in long-term academic development.
5. Application Challenges and Optimization Strategies
5.1 Practical Application Challenges of XAI Optimization Model
Although the XAI-optimized student behavior prediction model achieves good performance in empirical analysis, there are still some practical challenges in the popularization and application of actual higher education scenarios. First, the diversity and complexity of educational data bring interpretation difficulties. Higher education learning scenarios are diverse, and student learning behaviors are affected by professional characteristics, curriculum difficulty, personal learning foundation and other multiple factors. The heterogeneity of multi-scene and multi-group data makes the feature interpretation mechanism have certain scenario limitations, and the universal interpretation rules are difficult to form[15].
Second, the real-time performance of model interpretation needs to be improved. The current model realizes offline batch interpretation of historical data, but cannot complete real-time monitoring and dynamic interpretation of student learning behaviors in the teaching process. The lag of interpretation results affects the timeliness of teaching intervention and fails to realize real-time early warning and dynamic guidance of student learning risks.
Third, the integration of technical interpretation results and educational practice is insufficient. The XAI model outputs quantitative feature contribution data, but lacks targeted educational intervention suggestions matched with interpretation results. Most interpretation results stay in the data analysis level, and cannot be directly transformed into operable teaching optimization schemes and student learning improvement strategies, resulting in disconnection between technical analysis and educational practice.
Fourth, the model’s personalized adaptation ability is limited. The current model is optimized based on overall student group data, and cannot fully adapt to the learning behavioral characteristics of students of different majors, different grades and different learning foundations. The interpretation results have limited pertinence for heterogeneous student groups.
5.2 Targeted Optimization and Improvement Strategies
Aiming at the above practical application challenges, this study puts forward four targeted optimization strategies to further improve the practicability and universality of the XAI optimization model. First, build a scenario-adaptive interpretable framework. According to the differences of different majors, different curriculum types and different teaching scenarios, classify and optimize the feature contribution rules, construct grouped interpretation models for different student groups, eliminate the interference of scenario heterogeneity on interpretation results, and improve the universality of model interpretation.
Second, realize real-time dynamic interpretation and early warning. Integrate the model with the campus smart teaching platform, collect student learning behavioral data in real time, update the model feature contribution analysis results dynamically, realize real-time monitoring of student learning status and instant early warning of learning risks, and improve the timeliness and effectiveness of teaching intervention.
Third, build a closed-loop mechanism of interpretation results and educational intervention. Correlate the quantitative feature contribution results with educational intervention measures, formulate targeted teaching optimization schemes and student learning guidance strategies corresponding to different behavioral deficiencies, realize the transformation from data interpretation results to practical teaching actions, and solve the disconnection between technical analysis and educational practice.
Fourth, strengthen the personalized optimization ability of the model. Introduce personalized feature weight adjustment mechanism, dynamically adjust the contribution weight of behavioral features according to individual student learning foundation and learning characteristics, realize personalized prediction and interpretation for individual students, and further improve the fine-grained analytical ability of the model.
6. Research Conclusions and Future Prospects
This study takes the optimization of higher education learning analytics student behavior prediction model as the core goal, applies XAI technology to break the black-box dilemma of traditional prediction models, constructs a systematic hybrid optimization framework of LightGBM and dual XAI interpretable algorithms, and verifies the effectiveness and practicability of the model through empirical data analysis. The core research conclusions are summarized as follows.
First, the XAI optimization strategy can realize the dual improvement of model transparency and predictive stability. The optimized model maintains the high-precision predictive performance of the traditional machine learning model, and effectively solves the defects of opaque decision-making and uninterpretable results of the black-box model, realizing the organic unity of predictive accuracy and model interpretability.
Second, student learning task quality, online learning duration and classroom attendance state are the core behavioral factors affecting college students’ academic performance, which have decisive positive effects on student learning outcomes. Interactive participation behavior plays an auxiliary promoting role, and the comprehensive superposition of multiple behavioral factors determines the final academic level of students.
Third, the hybrid XAI interpretation framework combining SHAP and LIME can realize complementary global and local analysis. SHAP global interpretation clarifies the overall rule of behavioral feature contribution, and LIME local interpretation realizes personalized diagnosis of individual student learning problems, forming a multi-dimensional and full-coverage model interpretation system.
Fourth, the practical application of XAI optimization model in higher education still faces challenges such as scenario heterogeneity, insufficient real-time performance and disconnection between analysis and practice. The targeted optimization strategies proposed in this study can effectively solve the above problems and provide practical guidance for the popularization and application of interpretable prediction models in educational scenarios.
6.2 Research Deficiencies and Future Prospects
This study realizes the optimization and interpretation of student behavior prediction models in higher education learning analytics, but there are still some research deficiencies. First, the research data is limited to single university samples, and the universality of the model needs to be further verified by multi-school and multi-region data. Second, this study only focuses on conventional learning behavioral indicators, and does not fully cover implicit behavioral factors such as student learning motivation and learning psychology, which needs to be further expanded in subsequent research. Third, the model’s dynamic evolution ability for long-term student learning behavior changes needs to be further optimized.
Future research can be carried out from three aspects. First, expand the research data scope, collect learning data of different types of universities and different student groups, optimize the model framework, and improve the universality and generalization ability of the interpretable model. Second, integrate multi-modal data such as student psychological characteristics and cognitive level, enrich the feature indicator system, and realize more comprehensive and in-depth analysis of student learning behavior rules. Third, build a long-term dynamic prediction and interpretation model, track the dynamic evolution of student learning behaviors in the whole academic stage, realize continuous monitoring and cumulative analysis of student learning development, and provide more scientific and systematic decision-making support for long-term student training and higher education teaching optimization.
Reference
[1]Kushwaha S P ,Badhera U ,Kamila K M . Impact of global health emergency on learning analytics research in higher education: a bibliometric analysis[J].Journal of Science and Technology Policy Management,2026,17 (5):1084-1106.
[2]Celbert M H ,Blesie V ,Nelson F N , et al. Adoption of ChatGPT Among Higher Education Students: A PLS-SEM Analysis Using the UTAUT-2 Framework[J].International Journal of Information and Communication Technology Education (IJICTE),2026,22 (1):1-29.
[3]Qadri A H A ,Mouas S ,Nouari W , et al. Factors influencing blended learning acceptance in Algerian higher education: a comprehensive analysis[J].Library Hi Tech,2026,44 (1):41-63.
[4]Yüksel A ,Ömerustaoğlu A ,Filiz A , et al. Sustainable Digital Learning in Higher Education: Insights from Student Analytics and Participation in BirDeHa[J].Sustainability,2026,18 (10):4980-4980.
[5]Fan C ,Ke L ,Chen Z , et al. Exploring the effect of GenAI on learning outcomes in higher education: a three-level meta-analysis[J].Frontiers in Psychology,2026,17 1758670-1758670.
[6]Chen W ,Xu W . The effectiveness of technology-enhanced L2 learning in higher education on academic performance from a cognitive load perspective: A meta-analysis[J].System,2026,140 104068-104068.
[7]Słowiński G ,Kapranov Y ,Iwanowska B . Threshold-like Structure and Barrier–Benefit Asymmetry in Students’ Behavioural Intention Toward E-Learning: An Explainable Machine Learning Analysis in Post-Pandemic Higher Education[J].Sustainability,2026,18 (10):4633-4633.
[8]Nguyen P T . Graduate skill gaps in applied higher education: a triadic analysis from a work-based learning perspective[J].Higher Education, Skills and Work-Based Learning,2026,16 (3):677-692.
[9]Drăgoi V M ,Puiu A R ,Petrea G , et al. Willingness to Allow Educational Data Use for Learning Analytics in Higher Education: Trust and Governance Predictors: An Exploratory Study[J].Education Sciences,2026,16 (4):637-637.
[10]Sánchez O D ,González P C . Psychometric validation of the ‘ChatGPT acceptance and use scale’ (AUS-GPT), and comparative analysis of its adoption among Mexican higher education students (2023-2025)[J].Social Sciences & Humanities Open,2026,13 102717-102717.
[11]Erdem C ,Kaya M ,Toptaş T H , et al. Problem-based learning and student outcomes in higher education: a second-order meta-analysis[J].Studies in Higher Education,2026,51 (4):950-971.
[12]Gondaliya V . Analyzing the English E-Learning Environment at PAAET: Pedagogical and Technological Perspectives from Kuwaiti Higher Education[J].Journal of Research in Vocational Education,2026,8 (3):42-45.
[13]Agal S . A privacy preserving synthetic learner dataset for learning analytics in technology enhanced higher education.[J].Scientific reports,2026,16 (1):14772-14772.
[14]Kaya M ,Erdem C ,Güngör F , et al. Team-based learning improves student performance in tertiary education: A second-order meta-analysis[J].Innovations in Education and Teaching International,2026,63 (2):571-584.
[15]Alqurni J . Sustainable adoption of artificial intelligence and the Metaverse in higher education: an environmental, social, and governance–based analysis of pedagogical innovation and perceived student learning outcomes[J].Frontiers in Artificial Intelligence,2026,9 1738730-1738730.