Personalized learning resource recommendation is an essential component of intelligent tutoring systems. To address the issue of the plethora of learning resources and enhance the learner experience in intelligent tutoring systems, learning resource recommendation systems have been developed to model learners’ preferences. Despite numerous efforts and achievements in academia and industry toward more personalized learning, intelligent education tailored to individual learners still faces challenges, such as inadequate user representation and potential information loss during the aggregation of multi-source heterogeneous information features. In recent years, knowledge-graph-based recommendation systems have brought hope for mitigating these issues and achieving more accurate recommendations. In this paper, we propose a novel personalized learning resource recommendation method based on a knowledge graph named the Learner-Enhanced Knowledge Graph Attention (LKGA) network. This model enhances learner representation by extracting collaborative signals, where learning resources clicked by learners who have clicked the same resource are considered potential collaborative signals and are concatenated with the original learning resource features to form the initial entity set for the learner. Furthermore, during the entity aggregation process, each tail entity has different semantic expressions, and an attention mechanism is used to distinguish the importance of different neighbor entities. Additionally, residual connections are added in each hop of the learner’s aggregation process, with the information from the first hop added to each subsequent hop to reduce information loss. We applied the proposed LKGA model to a real-world dataset, and the experimental results fully validate the effectiveness of our model.

The rapid development of intelligent learning guidance systems has created a favorable environment for personalized learning. By accurately predicting students’ future performance, education can be tailored and teaching strategies optimized. However, traditional prediction algorithms seldom account for highly imbalanced datasets in basic education, overlook temporal factors, and lack further interpretability of the prediction results. To address these shortcomings, we propose Temporal Variational Autoencoder-Generative Adversarial Network (TVAE-GAN), a temporal variational autoencoder-generative adversarial network model aimed at providing early warnings for high-risk students in basic education, with in-depth interpretability analysis of the prediction results to suit the unique context of basic education. TVAE-GAN extracts features from real samples and introduces a Long Short-Term Memory (LSTM) network to capture dynamic features in time series, helping the model better understand temporal dependencies in the data, remember the sequential causal information of students’ online learning, and achieve better data generation performance. Using these features, the generative model generates new samples, and the discriminator model evaluates their quality, producing outputs that closely resemble real samples through training. The effectiveness of the TVAE-GAN model is validated on a collected online basic education dataset while also advancing the timing of interventions in predictions. The performance differences between the proposed method and classic resampling methods, as well as their impact in the educational field, are analyzed, highlighting that misclassification increases teacher workload and affects students’ emotions. Key influencing factors are identified using a decision-tree surrogate model, providing teachers with multidimensional references for academic assessment.

Online learning has become a mainstream instructional modality worldwide. However, early warning models suffer from three critical deficiencies in hard sample identification: temporal feature dependency, class imbalance, and weak feature extraction capability for boundary-ambiguous samples. This paper proposes an early warning framework, FocalGAN-SEE, which comprises a Focal Generative Adversarial Network (FocalGAN) that generates hard boundary samples to alleviate class imbalance, and an SE attention-enhanced encoder (SE Encoder) that adaptively amplifies discriminative features of hard at-risk samples through temporal channel recalibration. Unlike standard generative models, the FocalGAN incorporates a focal hard-sample loss that drives the generator to produce boundary-ambiguous samples, thereby enhancing the classifier’s robustness to hard samples. We compared the proposed architecture with attention-enhanced autoencoders, deep generative models, and oversampling-based models currently used in online learning early warning systems. The results show that the proposed FocalGAN-SEE achieves superior performance across accuracy, precision, recall, F1-score, and AUC, and can identify more hard at-risk samples in earlier course modules. Data visualization shows that the focal loss successfully guides the model to focus on the hard region with the highest overlap in the feature space, and FocalGAN-SEE can detect more hard samples at ambiguous boundaries. Finally, we adopt SHAP to interpret the prediction logic of the model, so as to assist teachers in implementing personalized teaching interventions. This paper provides a principled solution to hard at-risk sample recognition in imbalanced sequential data and offers practical value for online course management and personalized interventions.