Abstract
In order to solve the problems of inefficient allocation of teaching resources and inaccurate recommendation of learning paths in higher education, this paper proposes a smart education optimization model (SEOM) by combining the improved random forest algorithm (RFA) based on adaptive enhancement mechanism and the Graph Neural Network (GNN) algorithm. The public data and information such as the national higher education intelligent education platform are collected, and SEOM is trained and verified. The results show that SEOM has high accuracy and generalization ability in three different teaching scenes: online mixed teaching, personalized teaching and project-based teaching. The Root Mean Square Error (RMSE) value in cross-validation is between 0.2 and 0.5, and the Mean Absolute Error (MAE) value is between 0.1 and 0.5. SEOM shows strong stability when dealing with multidimensional educational resources and complex teaching modes. The accuracy rate remains at 85-97%, indicating its reliability in personalized learning path recommendation. Further analysis shows that the chi-square freedom ratio is between 1.0 and 2.5, the fitting index and the adjusted fitting index are both above 0.85, and the comparative fitting index is close to 0.95, which shows that SEOM has high accuracy and rationality in capturing the dependence of knowledge points in different teaching modes. The Root Mean Square Residual (RMR) and Root Mean Square Error of Approximation (RMSEA) are both below 0.05, which indicates that SEOM has small residual and strong scene adaptability. In addition, in the abnormal network environment, the resource allocation efficiency of SEOM is above 60%, and the Shapley value is between 0.1 and 0.4, which shows that SEOM can adapt to the change of network environment and the resource allocation effect is still obvious. Generally speaking, SEOM can optimize the allocation of educational resources and recommend learning paths in a complex environment, and effectively improve the intelligence and efficiency of teaching decision-making, especially for university administrators and educational technology developers.
Similar content being viewed by others
Introduction
Research background and motivations
In the era of digital intelligence, with the continuous integration of digital technology and intelligent technology, higher education is facing the need for profound changes1. Cutting-edge technologies such as cloud computing, big data, Internet of Things (IoT) and Artificial Intelligence (AI) drive the transformation of education from “digital” to “intelligent”. It not only promotes the automation process, but also requires the optimization of human-computer interaction and teaching mode to improve the quality and efficiency of personnel training2,3,4. In order to meet the social demand for innovative and compound talents, higher education urgently needs to rely on AI technology to build a personalized learning path and a smart education management system. Exploring the application of AI in higher education is not only an inevitable choice to adapt to the development of technology, but also an important path to optimize future talent training5.
Research objectives
The rapid development of AI is profoundly affecting all aspects of higher education. Traditional teaching mode, scattered arrangement of resources and single way of imparting knowledge have been difficult to adapt to the rapid development of emerging science and technology and the changes of social needs, and promote the accelerated transformation of higher education to intelligent, personalized and modular direction6,7. Based on this, this paper combines the improved Random Forest Algorithm (RFA) based on AdaBoost and the Graph Neural Network (GNN) algorithm to generate a smart education optimization model (SEOM). On the one hand, SEOM aims to improve the ability of feature extraction and classification of multidimensional learning resource data, and solve the problem of over-fitting of traditional algorithms on high-noise datasets. On the other hand, by introducing GNN, people can optimize the modeling ability of the complex relationship between students, learning resources and knowledge points, realize more accurate personalized learning path recommendation and resource allocation optimization, and improve the decision-making efficiency of the teaching system.
Literature review
In the field of the integration of higher education and AI, many scholars have conducted in-depth research to explore the potential of intelligent technology in teaching reform and resource optimization. Abdurashidova et al. studied the influence of digital technology on the quality of higher education in Tashkent State University of Technology, and investigated 300 participants with mixed methods. The results showed that 83% of the respondents believed that digital technology had a significant positive impact on the quality of education, and 72% of the respondents said that they could obtain useful digital resources. The research further emphasized the core role of digital technology in improving the quality of higher education and promoting dynamic learning environment8. Zheng et al. put forward the importance of AI technology in the construction of intelligent education platform, especially through the combination of evolutionary algorithm and deep belief network, and built a personalized curriculum recommendation model. The research showed that the performance of the model on public datasets was better than that of traditional methods, which showed the great application potential of machine learning in the intelligent education system9.
Meanwhile, Grajeda et al. took private universities in Latin America as the research object, and discussed the application and influence of AI tools in higher education. Through the investigation of 4127 students, the analysis showed that AI tools had significantly enhanced students’ understanding, creativity and productivity10. In addition, Badal and Sungkur focused on the application of online learning platform. Through the introduction of RF classifier, the accuracy rate reached 85%, which successfully predicted students’ performance and participation, demonstrating the superiority of machine learning in predicting students’ grades11. Lei et al. put forward a multi-agent learning model based on GNN, aiming at overcoming the limitation that the student agent can only get advice from a single teacher agent in the traditional teacher-student framework. It was found that by modeling the connection structure of agents, GNN had achieved significant performance improvement in learning strategy optimization, and its learning performance was superior to that of baseline method12. In the field of large-scale learning behavior data mining, Yuan et al. proposed an online course evaluation model based on graph automatic encoder. By introducing K-nearest neighbor graph to construct the association between online courses, and learning useful implicit features through variational graph automatic encoder, it was proved that it showed significant group coherence and structural differences in clustering tasks. In the classification task, the overall performance of the model was improved by about 10%, which fully verified its effectiveness in dealing with complex network data13. The above research shows that with the continuous development of digital technology and intelligent algorithms, the teaching mode, resource allocation and students’ learning behavior analysis of higher education have entered a new stage of intelligence and personalization.
Although previous studies have made remarkable progress in the application of intelligent education technology, most of them focus on the realization of a single algorithm or the optimization of local scenes. In addition, the dynamic adaptability of teaching resources under complex network conditions has not been effectively discussed. Based on this, this paper puts forward SEOM model by improving the fusion of RFA and GNN, aiming at improving the intelligent level of resource allocation in complex environment and providing new theoretical and technical support for the optimization of personalized teaching path in higher education.
Research model
The realization of SEOM can be summarized as follows: RFA uses a large number of educational resources samples to extract features efficiently through its integrated learning mechanism14,15. However, with the increase of the number of decision trees, RF show bottlenecks in space and time complexity, and their “black box” nature makes the model less interpretable, especially in noisy datasets16,17,18. In order to solve this problem, this paper improves the RFA by introducing AdaBoost. By improving the accuracy of weak classifiers through the weighting mechanism, AdaBoost can better capture the key features of multi-dimensional learning resources, making the model more accurate and stable in complex teaching scenes. However, it is difficult to fully reflect the complex interaction among students, knowledge points and teaching resources by relying solely on RF19,20. Therefore, SEOM further combined with GNN, using GNN to model the graph structure between students and knowledge points, and capturing the dependence between knowledge points through the update and dissemination of nodes and edges21. This method forms a dynamic learning path optimization mechanism, which is helpful to realize more accurate learning path recommendation and educational resource allocation optimization. The result of feature selection provided by RF is used as GNN input, which promotes the dynamic adjustment of graph structure and makes the model more intelligent to deal with complex educational scenes22,23,24. Finally, through continuous feedback and iteration, SEOM can improve the overall learning effect, realize efficient personalized learning path recommendation and optimal allocation of resources, and promote the intelligent reform of higher education. The implementation process of SEOM is shown in Fig. 1:
In this process, SEOM optimizes the logical structure of feature extraction and classification by combining AdaBoost with RFA. The multi-tree structure of RF comprehensively analyzes the feature distribution of high-dimensional educational data through parallel processing to identify potentially important variables, while AdaBoost gradually optimizes the error of classifier based on iterative adjustment mechanism, thus improving the robustness of the model in multiple rounds of training. By updating the dynamic weights, AdaBoost assigns more attention to the samples that are difficult to classify, so that the model can filter out useful features more effectively when dealing with noisy data. In addition, this method also enhances the comprehensive utilization of weak classifiers by RF, and further improves the depth of feature selection and the accuracy of classification results. In the concrete implementation, SEOM ensures that each round of optimization can accurately reflect the important characteristics of educational data by introducing Gini gain and information entropy, which provides a more efficient modeling basis for learning path optimization and resource allocation. This design logic makes the model have strong adaptability and flexibility when dealing with complex educational scenes.
In the process of SEOM implementation, firstly, the RFA is applied to process a large number of high-dimensional student data, and the purity of the data is measured by the information entropy equation:
\(\:H\left(X\right)\) represents the information entropy of dataset X. \(\:p\left({x}_{i}\right)\) is the probability of sample \(\:{x}_{i}\) in the dataset, which is used to calculate the uncertainty of classification25. In order to further select the best segmentation point, Gini coefficient is introduced to measure the purity of nodes:
\(\:G\left(X\right)\) is the Gini coefficient of characteristic \(\:X\), and the optimal splitting is selected by reducing impurity. In order to improve the performance of the model, AdaBoost is introduced to adjust the learning ability of the model through the sample weighting mechanism. Its weight updating equation is:
\(\:{w}_{i}^{(t+1)}\) represents the weight of sample \(\:i\) after the \(\:t+1\) iteration. \(\:{y}_{i}\) is the actual label of sample \(\:i\). \(\:{f}_{t}\left({x}_{i}\right)\) is the prediction result of weak classifier for \(\:{x}_{i}\). \(\:{\alpha\:}_{t}\) is the weight coefficient based on the classifier error rate, which is calculated as follows:
\(\:{\epsilon}_{m}\) is the error rate of the \(\:m\)-th weak classifier. \(\:{\alpha\:}_{m}\) is used to weight the contribution of the weak classifier. Through this weighting mechanism, the model gradually improves the classification accuracy26,27. Meanwhile, when the nodes are split, the feature selection is optimized through the Gini gain calculation equation:
\(\:{\Delta\:}G\) represents Gini gain. \(\:{X}_{k}\) is the \(\:k\)-th subset. \(\:\left|{X}_{k}\right|\) is its sample size, and \(\:\left|X\right|\) is the sample size of the parent node. This equation selects the optimal splitting point by maximizing Gini gain.
In order to better model the complex learning path and the relationship between students and knowledge points, RF is combined with GNN. GNN updates the node characteristics of each layer by the following equation:
\(\:{h}_{i}^{(l+1)}\) is the eigenvector of node \(\:i\) in the \(\:l+1\) layer. \(\:N\left(i\right)\) represents the neighbor node set of node \(\:i\). \(\:{d}_{i}\) and \(\:{d}_{j}\) are the degrees of node \(\:i\) and \(\:j\) respectively, and \(\:\sigma\:\) is the activation function. GNN uses the information dissemination mechanism between nodes and edges to update the weight of edges between nodes through the following equation:
\(\:{e}_{ij}\) is the weight of the edge between nodes \(\:i\) and \(\:j\). \(\:\overrightarrow{a}\) is the learning vector in the attention mechanism. \(\:W\) is the learnable weight matrix, and \(\:\parallel\:\) represents the splicing operation of vectors28. Through this process, the system can capture the complex dependence between students and knowledge points, and then optimize the learning path29.
The final prediction output of the whole model is determined by the weighted fusion of several weak classifiers, and the equation is as follows:
\(\:F\left(x\right)\) is the final classification result. \(\:{\alpha\:}_{m}\) is the weight of each classifier. \(\:{f}_{m}\left(x\right)\) is the prediction result of sample \(\:x\) by the \(\:m\)-th classifier30. In order to prevent over-fitting, the expected error equation is introduced into the model:
\(\:\text{E}\left[\text{E}\text{r}\text{r}\right(h\left)\right]\) represents the expected value of the error rate. \(\:\mathbb{I}\left(h\right({x}_{i})\ne\:{y}_{i})\) is the indicator function, and it is judged whether the predicted \(\:h\left({x}_{i}\right)\) is inconsistent with the real label \(\:{y}_{i}\).
In GNN, the cross-entropy loss function is also introduced to measure the classification effect of the model:
\(\:\mathcal{L}\) is the loss function. \(\:{y}_{i}\) is the true label of sample \(\:i\). \(\:\widehat{{y}_{i}}\) is the prediction probability of the model. In order to further optimize the node features, GNN uses the following feature aggregation equation:
\(\:\mathcal{N}\left(i\right)\) is the neighbor node of node \(\:i\). \(\:W\) is the weight matrix. The features of the current node are updated by aggregating the features of the neighbor nodes31.
In the whole optimization process, the contribution of each feature is measured by the calculation equation of feature importance:
\(\:\text{Feature\:Importance}\left({x}_{j}\right)\) indicates the importance of feature \(\:{x}_{j}\). \(\:{w}_{i}\) is the sample weight, and \(\:G\left({x}_{j}\right)\) is the Gini gain of feature \(\:{x}_{j}\)32,33.
GNN updates the feature information between nodes through the messaging mechanism, and the equation is as follows:
\(\:{m}_{ij}\) is the information transmitted from node \(\:j\) to node \(\:i\). \(\:{h}_{j}^{\left(l\right)}\) is the feature vector of node \(\:j\) at the \(\:l\) layer.
Finally, the equation for optimizing the learning path is:
\(\:{P}_{\text{optimized}}\left(x\right)\) is the optimized learning path. \(\:P\left({x}_{i}\right)\) is the initial estimate of the path. In order to ensure the effective learning of the model in the training process, the back propagation mechanism is used to update the weight matrix, and the equation is as follows:
\(\:\frac{\partial\:\mathcal{L}}{\partial\:W}\) represents the gradient of the loss function to the weight \(\:W\). \(\:\frac{\partial\:\mathcal{L}}{\partial\:{h}_{i}^{(l+1)}}\) is the gradient of the node characteristics. \(\:{h}_{i}^{\left(l\right)}\) is the node characteristics of the upper layer.
Through multiple rounds of optimization and feedback of RFA and GNN algorithm, the prediction accuracy of SEOM and the optimization effect of learning path are continuously improved34.
Experimental design and performance evaluation
Datasets collection
The data studied in this paper mainly comes from open educational platforms and databases, covering three types of datasets. Firstly, the dataset of educational resources is collected through MOOC platform and the national platform of higher education wisdom education. After cleaning, more than 11,000 records of online courses, teaching materials and teaching videos are kept. Secondly, the teaching model datasets of different majors come from China Higher Education Student Information Network, which covers the teaching plans and curriculum structure data of more than 200 majors in colleges and universities across the country. After standardization, they are combined for model analysis35. In addition, due to the limited access to the characteristic data of students’ behavior teaching, 8,000 students’ learning behavior data are captured and simulated through the teaching records made public by local ministries of education, and combined with information such as study habits, attendance rate and grades, a multi-dimensional student learning dataset is formed. After data cleaning, 37,892 valid records are finally retained, and the variables with correlation coefficient greater than 0.3 are screened by correlation analysis in the feature extraction stage. After data preprocessing, 28,6062 data are retained for model training and testing.
The data information of training and testing is shown in Table 1:
Experimental environment
The training and testing of SEOM model is based on the core goal of model design, and its performance is comprehensively and objectively tested from multiple levels and angles. Firstly, in terms of the accuracy and generalization ability of the model, cross-validation (K = 10) and set-aside validation are adopted to test its ability to maintain high prediction accuracy and strong generalization under different data distributions36. Meanwhile, based on Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), the error of the model is quantified to verify its ability to deal with multi-dimensional educational resources and complex teaching models. Secondly, in the aspect of learning path optimization, by calculating Cmin/df, Goodness of Fit Index (GFI), Adjusted Goodness of Fit Index (AGFI), Comparative Fit Index (CFI), Root Mean Square Residual (RMR) and Approximate Root Mean Square Error (RMSEA), and analyze the recommendation effect of the model on personalized learning paths in different major categories (engineering, science, medicine, law, education, management, literature, economics, art and agriculture) in common teaching scenes37,38,39. In addition, by setting an abnormal network environment, Shapley value is used to analyze the contribution and delay sensitivity of resource allocation, and the intelligent effect of resource allocation in complex environment is tested40,41.
Among them, the evaluation criteria of each index in the analysis of learning path optimization effect are shown in Table 2:
Parameters setting
In this paper, the Graph Attention Network (GAT) is selected as the core type of GNN, and the features of neighboring nodes are given different weights through the attention mechanism to capture the uneven dependence between nodes more accurately. GAT is especially suitable for personalized learning path planning and educational resource optimization because of its strong expressive ability and adaptability to complex relationships between nodes. GAT is used to model the multi-level interaction between students and knowledge points, dynamically adjust the weights of adjacent nodes through attention mechanism, and generate a feature vector representation that is more suitable for the actual learning scene. Its architecture consists of three layers of attention, and each layer adopts a multi-head attention mechanism to improve the robustness and global nature of feature capture. In the application environment, GAT first receives the high-quality features screened by RFA as input and constructs the initial graph structure. Then, through iterative feature updating and aggregation, the dependency map of learning path is gradually optimized, and finally personalized recommendation results are generated. The main reason for choosing GAT is that it can effectively reduce the over-smoothing problem compared with the traditional GNN, especially in the dynamic education scene, and it can handle the complex interaction between students’ learning behavior and knowledge points more accurately through the propagation of node characteristics and the weighted update of edges. SEOM hyperparameters and training methods are arranged as shown in Table 3:
Relevant parameters of common teaching scenes in the analysis and verification of model learning path optimization effect are shown in Table 4:
The relevant parameters of sudden abnormal network environment in the analysis and verification of model educational resource allocation effect are shown in Table 5:
Of course, in order to better highlight the advantages of SEOM in this paper, referring to the literature review in the previous paper, five baseline models that are most relevant to this paper are selected, compared and verified under the same data standard (10,000 data in the dataset are randomly selected). The selected baseline models are sorted out, and the results are shown in Table 6:
The software and hardware environment parameters in the whole training and testing of the model are shown in Table 7:
Performance evaluation
Analysis of accuracy and generalization ability of the model
The analysis results of SEOM model accuracy and generalization ability are shown in Fig. 2:
In Fig. 2, SEOM model shows high accuracy and good generalization ability in cross-validation (K = 10) and set-aside validation. The RMSE value of cross-validation fluctuates between 0.2 and 0.5, and the MAE value is between 0.1 and 0.5, which shows the stability of the model in dealing with multi-dimensional educational resources and complex teaching models. Meanwhile, the accuracy of the model remains at 85-97%, indicating its reliability in optimizing educational resources and recommending learning paths. The comparison between RMSE and MAE is slightly higher, which keeps fluctuating between 0.2 and 0.6, but within the accuracy range of 80-90%, indicating that the model still has strong adaptability and wide application potential.
Effect analysis of model learning path optimization
The analysis result of SEOM model learning path optimization effect is shown in Fig. 3:
Figure 3 shows that the Cmin/df of SEOM model is between 1.0 and 2.5 in different major categories and teaching scenes, indicating that the model has an ideal fitting degree in each scene. The indexes of GFI and AGFI are above 0.85, and CFI is close to 0.95, which shows that the model has high accuracy and rationality in capturing the dependence of knowledge points in different teaching modes. Meanwhile, RMR and RMSEA values are lower than 0.05, which shows that the residual of the model is small and the model has strong adaptability to the actual teaching scene. Especially in the personalized teaching scene, the recommendation effect of the model on the learning path is significantly improved, which further verifies SEOM’s intelligent optimization ability in the complex educational environment.
Effect analysis of model education resource allocation
The analysis results of educational resource allocation efficiency of SEOM model are shown in Fig. 4:
In Fig. 4, the Shapley value ranges from 0.1 to 0.4 in the abnormal network environment, indicating that there are significant differences in the contribution of different network environments to resource allocation. Among them, bandwidth bottleneck and network delay have great influence on the efficiency of resource allocation, and the delay sensitivity is high, up to 0.8, which shows that network delay has a strong interference effect on the allocation of teaching resources. The efficiency of resource allocation fluctuates between 60% and 95%, which depends on the stability of the network environment. The analysis results show that SEOM can intelligently optimize the allocation of educational resources under complex network conditions, especially in the environment of high delay and bandwidth bottleneck, and the model shows strong adaptability and resource allocation ability to ensure the effective completion of teaching tasks.
Comparative analysis of model with other baseline models
The comparative analysis results of SEOM model and other baseline models are shown in Fig. 5:
As Fig. 5 shows, SEOM model has obvious advantages in many key indicators, especially when dealing with multi-dimensional educational resources optimization and personalized learning path recommendation, and its performance is superior to the traditional baseline model. In terms of model accuracy, the accuracy of SEOM is always above 95% when the data scale is gradually increased to 10,000, which is significantly higher than the stable values of GCN and SVM models of 88% and 83%, showing excellent generalization ability and robustness. At the same time, in the modeling of complex dependencies, SEOM effectively optimizes the distribution of feature weights by virtue of the attention mechanism of GATT. The RMR value is reduced to 0.035, which is significantly lower than the average value of 0.052 in other baseline models, reflecting that the fitting error of the model to multidimensional data is significantly reduced. In addition, in the optimization of resource allocation efficiency and learning path, SEOM maintains the allocation efficiency of more than 90% in the high data density environment through the dynamic weighted adjustment of Shapley value. Additionally, its resource adaptation ability is more prominent than that of KNN model, where the efficiency drops to 75%. This shows that SEOM optimizes the complex dependence among knowledge points in the personalized learning scene, and effectively realizes the intelligent allocation of teaching resources in the complex environment. This fully reflects its wide application potential in the intelligent reform of higher education.
Discussion
SEOM model combines RF, AdaBoost and GNN to build an intelligent optimization system for complex educational resources and personalized learning paths. RF improves the accuracy of multi-dimensional educational resources processing by means of adaptive enhancement mechanism, and GNN enhances the accuracy of the model in learning path prediction by constructing the dependency map between students and knowledge points. The model verification results show that the fitting indexes such as Cmin/df, GFI, AGFI and CFI are highly accurate, and the indexes of RMR and RMSEA are in a reasonable range, which indicates that SEOM has high fitting ability and small error. Especially in the abnormal network environment, SEOM shows excellent resource allocation optimization ability through Shapley value and delay sensitivity analysis. In addition, SEOM can intelligently recommend the optimal learning path according to students’ learning habits and knowledge points in personalized learning scenes, and realize the seamless docking of personalized teaching through real-time data update. In the aspect of educational resource management, SEOM intelligently allocates and transmits all kinds of resources by integrating with the management platform to ensure the efficient allocation of resources and the smooth progress of teaching tasks.
From the optimization point of view, SEOM effectively alleviates the bottleneck problem of feature extraction in traditional education model when dealing with high-dimensional and complex data through the deep combination of improved RF algorithm and adaptive enhancement mechanism. RF algorithm improves the robustness of feature selection through multi-tree structure, and adaptive enhancement mechanism further strengthens the comprehensive performance of weak classifiers, enabling SEOM to accurately capture key features in educational resources and optimize resource allocation strategies. On this basis, GNN, by virtue of its modeling ability of dynamic dependency, not only comprehensively depicts the complex interaction relationship, but also dynamically adjusts the weight of nodes by constructing the map structure between students and knowledge points. It effectively copes with the problem of nonlinear feature distribution and dynamic change in learning path recommendation. The verification results further illustrate this point. SEOM significantly reduces the error rate (RMSE value is reduced to 0.2–0.5) in high-noise data environment. Its efficient fitting ability is verified by Cmin/df and GFI, especially in complex network scenarios, which can accurately optimize the resource allocation efficiency through Shapley value and delay sensitivity analysis. Compared with the traditional model, the architecture design of SEOM has shown outstanding advantages in intelligent resource allocation, personalized learning path recommendation and adaptability of teaching scenarios, thus promoting the comprehensive realization of intelligent optimization of higher education.
Conclusion
Research contribution
The innovation of this paper is mainly reflected in two aspects. Firstly, the multi-level algorithm fusion of model design shows strong adaptability and flexibility for large-scale dynamic data processing and personalized learning path recommendation, which solves the limitation that traditional single algorithm is difficult to deal with multi-dimensional dynamic data. By combining the fine feature selection of RF with GNN’s graph structure modeling ability, SEOM has realized the deep analysis of the dependence between students’ learning behavior and knowledge points, especially in capturing complex dependence and adjusting learning paths in real time, ensuring the accuracy of path recommendation in different teaching scenes. This design breaks through the static analysis limitation of traditional personalized recommendation system. Secondly, by combining learning path optimization and resource scheduling, SEOM puts forward a dynamic resource allocation framework driven by teaching tasks, which improves resource utilization and personalized learning experience. This closely integrated design optimizes educational resource management and path recommendation, and provides new impetus for the development of intelligent education system.
Future works and research limitations
Although SEOM has obvious advantages, there are still some details that need to be further improved. Firstly, the computational complexity of the model is high, especially in the face of large-scale and multi-dimensional data, which may increase the consumption of computing resources and affect the real-time processing ability. Future research can focus on optimizing the computational efficiency of the model and improving its performance in large-scale application scenes. Secondly, although the verification process of the model includes complex scene simulation, the diversity of the actual teaching environment is not fully reflected. Further research should introduce more actual data for verification to enhance the universality and stability of the model in different environments and provide more innovative algorithm framework for the field of educational technology.
Meanwhile, from the wider perspective of SEOM education promotion, with the continuous expansion of its application scope, the model may face more obvious scalability challenges when dealing with ultra-large-scale educational networks and multi-source heterogeneous data. Firstly, with the exponential growth of the number of nodes and edges, the computing and storage requirements of GNN show nonlinear growth, which may lead to a significant increase in system delay, thus weakening the real-time and response ability of SEOM in large-scale educational scenarios. In addition, the differences of technical basis in different geographical regions and educational environments may further lead to the problems of data inconsistency and system adaptability, and limit the wide applicability of the model. In order to solve these problems, future research should further optimize semantic modeling and multimodal fusion technology, and combine self-supervised learning with deep causal reasoning to intelligently optimize the whole link from perception to decision. By introducing dynamic knowledge map and situational awareness mechanism, SEOM’s adaptability and robustness in complex dynamic educational scenes are enhanced, so that it can accurately process multi-dimensional educational data and dynamically adjust resource allocation and learning path. In addition, the paper needs to focus on exploring the deep coupling between AIGC and emerging technologies such as metauniverse and immersive teaching. Moreover, the paper should create a seamless connection between virtual and real educational environment, give educational intelligence a broader development space, and provide intelligent solutions with deep learning and reasoning capabilities for future educational systems.
Data availability
The datasets used and/or analyzed during the current study are available from the corresponding author Lifeng Chen on reasonable request via e-mail chenlifeng@wzut.edu.cn.
References
Hetmanczyk, P. Digitalization and its impact on labour market and education. Selected aspects. Educ. Inform. Technol. 29 (9), 11119–11134 (2024).
Chaka, C. Fourth industrial revolution—a review of applications, prospects, and challenges for artificial intelligence, robotics and blockchain in higher education. Res. Pract. Technol. Enhanced Learn. 18 (7), 002–002 (2023).
Kamalov, F., Santandreu Calonge, D. & Gurrib, I. New era of artificial intelligence in education: Towards a sustainable multifaceted revolution. Sustainability 15(16), 12451 (2023).
Pan, L. et al. The impact of digital competence and personal innovativeness on the learning behavior of students: Exploring the moderating role of digitalization in Higher Education Quality. SAGE Open. 14(3), 21–58 (2024).
Crompton, H. & Burke, D. Artificial intelligence in higher education: The state of the field. Int. J. Educ. Technol. High. Educ. 20(1), 22 (2023).
Amnuaysin, O., Nilsook, P. & Wannapiroon, P. Enabling high performance digital manpower through higher education. World Trans. Eng. Technol. Educ. 21 (1), 38–43 (2023).
Qolamani, K. I. B. & Mohammed, M. M. The digital revolution in higher education: Transforming teaching and learning. QALAMUNA: Jurnal Pendidikan, Sosial, dan Agama 15(2), 837–846 (2023).
Abdurashidova, M. et al. The impact of innovation and digitalization on the quality of higher education: A study of selected universities in Uzbekistan. J. Intell. Syst. 32(1), 20230070 (2023).
Zheng, L. et al. Evolutionary machine learning builds smart education big data platform: Data-driven higher education. Appl. Soft Comput. 136(11), 110114 (2023).
Grajeda, A. et al. Assessing student-perceived impact of using artificial intelligence tools: Construction of a synthetic index of application in higher education. Cogent Educ. 11(1), 2287917 (2024).
Badal, Y. T. & Sungkur, R. K. Predictive modelling and analytics of students’ grades using machine learning algorithms. Educ. Inform. Technol. 28 (3), 3027–3057 (2023).
Lei, Y. et al. A GNN-based teacher–student framework with multi-advice. Expert Syst. Appl. 250 (6), 123887 (2024).
Yuan, W. et al. Online course evaluation model based on graph auto-encoder. Intell. Data Anal. 4 (4), 1–23 (2024).
Chew, Y. C. & Zainal, S. R. M. Revolutionising talent management in Society 5.0: Empowering organisations through collaborative talent management in the intelligence era. J. Electr. Syst. 20(4s), 1306–1321 (2024).
Alharbi, W. Mind the gap, please! Addressing the mismatch between teacher awareness and student AI adoption in higher education. IJCALLT 14 (1), 1–28 (2024).
Duah, J. E. & McGivern, P. How generative artificial intelligence has blurred notions of authorial identity and academic norms in higher education, necessitating clear university usage policies. Int. J. Inform. Learn. Technol. 41 (2), 180–193 (2024).
Lakshmi, A. J. et al. Artificial intelligence in steering the digital transformation of collaborative technical education. J. High. Technol. Manag. Res. 34 (2), 100467 (2023).
Ülku, A. Artificial intelligence-based large language models and integrity of exams and assignments in higher education: The case of tourism courses. Tourism Manag. Stud. 19 (4), 21–34 (2023).
Kim, J. Leading teachers’ perspective on teacher-AI collaboration in education. Educ. Inform. Technol. 29 (7), 8693–8724 (2024).
Southworth, J. et al. Developing a model for AI across the curriculum: Transforming the higher education landscape via innovation in AI literacy. Computers Education: Artif. Intell. 4 (6), 100127 (2023).
Nachouki, M. et al. Student course grade prediction using the random forest algorithm: Analysis of predictors’ importance. Trends Neurosci. Educ. 15 (4), 100214 (2023).
Praba, B. et al. Analysis of teaching-learning efficiency using attribute based double bounded rough neutrosophic set driven random forests. Neutrosophic Sets Syst. 55 (7), 13–34 (2023).
Choi, I. et al. Automatic scoring of mathematics descriptive assessment using random forest algorithm. Math. Educ. 63 (2), 165–186 (2024).
Alruwais, N. & Zakariah, M. Student-engagement detection in classroom using machine learning algorithm. Electronics 12 (3), 731 (2023).
Shi, Y. et al. Analysis of learning behavior characteristics and prediction of learning effect for improving college students’ information literacy based on machine learning. IEEE Access. 11 (16), 50447–50461 (2023).
Albahli, S. Efficient hyperparameter tuning for predicting student performance with bayesian optimization. Multimed. Tools Appl. 83 (17), 52711–52735 (2024).
Buyrukoğlu, S. A novel color labeled student modeling approach using e-learning activities for data mining. Univ. Access Inf. Soc. 22 (2), 569–579 (2023).
Almufarreh, A., Noaman, K. M. & Saeed, M. N. Academic teaching quality framework and performance evaluation using machine learning. Appl. Sci. 13 (5), 3121 (2023).
Chi, Z., Zhang, S. & Shi, L. Analysis and prediction of MOOC learners’ dropout behavior. Appl. Sci. 13 (2), 1068 (2023).
Pu, J. et al. Graph neural network based intelligent tutoring system: A survey. Neurocomputing 5 (3), 128442 (2024).
Caspari-Sadeghi, S. Artificial intelligence in technology-enhanced assessment: A survey of machine learning. J. Educational Technol. Syst. 51 (3), 372–386 (2023).
Bui, T. C., Le, V. D. & Li, W. S. Generating real-time explanations for GNNs via multiple specialty learners and online knowledge distillation. IEEE Access. 11 (5), 40790–40808 (2023).
Bukartaite, R. & Hooper, D. Automation, artificial intelligence and future skills needs: An Irish perspective. Eur. J. Train. Dev. 47 (10), 163–185 (2023).
França, T. J. F. et al. Artificial intelligence applied to potential assessment and talent identification in an organisational context. Heliyon 9 (4), 110–121 (2023).
Karimi, H. & Khawaja, S. The impact of Artificial Intelligence on higher education in England. Creative Educ. 14 (12), 2405–2415 (2023).
Alotaibi, N. S. & Alshehri, A. H. Prospers and obstacles in using artificial intelligence in Saudi Arabia higher education institutions—the potential of AI-based learning outcomes. Sustainability 15 (13), 10723 (2023).
Hannan, S. A. Development of digital transformation in higher education institutions. J. Comput. Sci. Comput. Math. 13 (01), 1–8 (2023).
Fernández, A. et al. Digital transformation initiatives in higher education institutions: A multivocal literature review. Educ. Inform. Technol. 28 (10), 12351–12382 (2023).
Jakubik, M. Cultivating the future in Higher Education: Fostering students’ life-world becoming with Wisdom Pedagogy. Trends High. Educ. 2 (1), 45–61 (2023).
Liu, C. GRA method for probabilistic simplified neutrosophic MADM and application to talent training quality evaluation of segmented education. J. Intell. Fuzzy Syst. 44 (5), 8637–8647 (2023).
Caspari-Sadeghi, S. Learning assessment in the age of big data: Learning analytics in higher education. Cogent Educ. 10 (1), 2162697 (2023).
Gollapalli, M. et al. SUNFIT: A machine learning-based sustainable University Field Training Framework for Higher Education. Sustainability 15 (10), 8057 (2023).
Siram, J., Chikati Srinu, D. S. V. & Tripathi, A. Towards a framework for performance management and machine learning in a higher education institution. J. Inf. Educ. Res. 4 (2), 32–41 (2024).
Chiu, T. & K F. Future research recommendations for transforming higher education with generative AI. Comput. Educ. Artif. Intell. 6 (8), 100197 (2024).
Funding
This work was supported by the Zhejiang Provincial Philosophy and Social Science Planning Project (No. 24SSHZ015YB); the Hangzhou Philosophy and Social Science Planning Project (No. 24JD055).
Author information
Authors and Affiliations
Contributions
Limin Qian: Conceptualization, methodology, software, validation, formal analysis, investigation, resources, data curation, writing—original draft preparation Weiran Cao: writing—review and editing, visualization, supervision, project administration, funding acquisition Lifeng Chen: writing—review and editing, visualization, supervision, project administration, funding acquisition.
Corresponding authors
Ethics declarations
Competing interests
The authors declare no competing interests.
Ethical approval
This article does not contain any studies with human participants or animals performed by any of the authors. All methods were performed in accordance with relevant guidelines and regulations.
Additional information
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
About this article
Cite this article
Qian, L., Cao, W. & Chen, L. Influence of artificial intelligence on higher education reform and talent cultivation in the digital intelligence era. Sci Rep 15, 6047 (2025). https://doi.org/10.1038/s41598-025-89392-4
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41598-025-89392-4