The continuous expansion of digital learning environments has catalyzed the demand for intelligent systems capable of providing personalized educational content. While current exercise recommendation frameworks have made significant strides, they frequently encounter obstacles regarding the long-tailed distribution of student engagement and the failure to adapt to idiosyncratic learning trajectories. We present LiveGraph, a novel active-structure neural re-ranking framework designed to overcome these limitations. Our approach utilizes a graph-based representation enhancement strategy to bridge the information gap between active and inactive students while integrating a dynamic re-ranking mechanism to foster content diversity. By prioritizing the structural relationships within learning histories, the proposed model effectively balances recommendation precision with pedagogical variety. Comprehensive experimental evaluations conducted on multiple real-world datasets demonstrate that LiveGraph surpasses contemporary baselines in both predictive accuracy and the breadth of exercise diversity.
Figures & tables
Figure 1: Overview of the LiveGraph framework for active-structure neural exercise recommendation. The Graph-Aware Student Representation Enhancer (Graph-VAE) transforms interaction history Ds into a stochastic mastery distribution θs , regularized by LLM-informed priors. In the Uncertainty-Aware Neural Re-ranker , candidate exercises C are scored based on a multi-signal fusion of relevance ϕrel , diversity ϕdiv , and Bernoulli Sub-graph Entropy U(e) . This fusion is dynamically orchestrated by a Meta-RL Controller that optimizes the exploration-exploitation trade-off via MAML -based adaptation. The Active Learning Probe identifies concept pairs with maximum mutual information I^(sij;Rij) to inject contrastive probes, triggering real-time Synchronous Kernel Evolution of the Dynamic Knowledge Kernel S(t) for continuous structural refinement.
Dataset
Students
Knowledge concepts
Exercises
Interactions
Nips34 [ 67 ]
4,918
57
984
1,399,470
Assist2009 [ 68 ]
4,217
123
17,737
374,422
Assist2012
27,066
265
45,716
2,541,201
Table 1: Dataset statistics. Each dataset consists of students, annotated knowledge concepts, exercises and student–exercise interactions.
Dataset
Model
NDCG@1
NDCG@3
NDCG@5
NDCG@10
F1@1
F1@3
F1@5
F1@10
Recall@1
Recall@3
Recall@5
Recall@10
Nips34 [ 67 ]
w/ All
0.942
0.967
0.961
0.949
0.942
0.908
0.861
0.765
0.942
0.881
0.808
0.682
w/o Graph-VAE
0.903
0.926
0.920
0.905
0.903
0.866
0.822
0.724
0.903
0.828
0.760
0.630
w/o Uncertainty
0.890
0.915
0.909
0.894
0.890
0.853
0.810
0.712
0.890
0.815
0.747
0.617
w/o Active Probe
0.878
0.901
0.896
0.882
0.878
0.841
0.799
0.701
0.878
0.802
0.735
0.604
w/o Meta-RL
0.865
0.888
0.883
0.870
0.865
0.828
0.786
0.690
0.865
0.788
0.722
0.592
Assist2009 [ 68 ]
w/ All
0.609
0.754
0.794
0.821
0.609
0.524
0.517
0.422
0.609
0.478
0.424
0.363
Table 2: LiveGraph Ablation Study: Component-wise analysis of model performance. "w/ All" represents the complete LiveGraph framework, while other configurations systematically remove key components.
Dataset
Model
NDCG@1
NDCG@3
NDCG@5
NDCG@10
F1@1
F1@3
F1@5
F1@10
Recall@1
Recall@3
Recall@5
Recall@10
Nips34 [ 67 ]
DKTRec [ 67 ]
0.697
0.822
0.830
0.824
0.697
0.736
0.726
0.688
0.697
0.650
0.620
0.567
DKVMNRec [ 69 ]
0.689
0.826
0.830
0.825
0.689
0.738
0.727
0.689
0.689
0.652
0.623
0.567
AKTRec [ 70 ]
0.740
0.853
0.855
0.846
0.740
0.768
0.749
0.700
0.740
0.688
0.648
0.582
SimpleKTRec [ 71 ]
0.735
0.852
0.854
0.845
0.735
0.772
0.749
0.702
0.735
0.692
0.650
0.584
KCPER [ 72 ]
0.338
0.545
0.609
0.646
0.338
0.488
0.541
0.590
0.338
0.401
0.432
0.465
KG4Ex [ 24 ]
0.486
0.693
0.731
0.732
0.486
0.590
0.608
0.622
0.486
0.486
0.487
0.495
Table 3: Comparative experiments across three datasets.
Dataset
Model
Group
NDCG@1
NDCG@3
NDCG@5
NDCG@10
F1@1
F1@3
F1@5
F1@10
Recall@1
Recall@3
Recall@5
Recall@10
Nips34 [ 67 ]
NR4DER-p [ 46 ]
Overall
0.915
0.939
0.934
0.919
0.915
0.879
0.828
0.727
0.915
0.839
0.762
0.622
LiveGraph
Overall
0.942
0.967
0.961
0.949
0.942
0.908
0.861
0.765
0.942
0.881
0.808
0.682
NR4DER-p [ 46 ]
Active
0.940
0.946
0.942
0.935
0.940
0.911
0.871
0.757
0.940
0.889
0.831
0.675
LiveGraph
Active
0.968
0.974
0.970
0.963
0.968
0.939
0.899
0.785
0.968
0.917
0.859
0.703
NR4DER-p [ 46 ]
Inactive
0.880
0.918
0.917
0.901
0.880
0.821
0.779
0.716
0.880
0.759
0.690
0.601
LiveGraph
Inactive
0.908
0.946
0.945
0.929
0.908
0.849
0.807
0.744
0.908
0.787
0.718
0.629
Table 4: Detailed performance analysis: LiveGraph vs NR4DER across different student groups and recommendation lengths.
Figure 2: Convergence behaviour: training steps vs. validation NDCG@5. Curves compare runs with and without the Meta-RL controller. Curves and shaded bands show mean ±1 standard deviation over five random seeds.
Figure 3: Diversity comparison across methods. Grouped bars show DIV@1, DIV@3, DIV@5 and DIV@10 for representative KT-based baselines, MMER/KCPER/KG4Ex, NR4DER-p (current SOTA), and LiveGraph. Bar heights correspond to the mean across seeds and error bars indicate one standard deviation.
Figure 4: Knowledge–concept distribution for a sampled student (Assist2009). The horizontal axis lists the 20 recommended exercises (sampled), the vertical axis lists the history frequency (top row) followed by 18 knowledge concepts. Color intensity denotes the presence/strength of a concept in each exercise (binary or continuous).
Figure 5: Hyperparameter sensitivity of the VAE prior coefficient β on validation NDCG@5 (log-scale x-axis). Each marker shows the mean validation NDCG@5 across seeds and vertical bars denote one standard deviation. The plot highlights the stable operating range used in our experiments.
(λ1,λ2)
NDCG@1
NDCG@3
NDCG@5
NDCG@10
Δ
(0.001,0.01)
0.597
0.747
0.786
0.814
−0.011
(0.01,0.1)
0.609
0.754
0.794
0.821
0.000
(0.1,1)
0.602
0.748
0.789
0.816
−0.006
(1,10)
0.592
0.740
0.780
0.807
−0.017
Table 5: Effect of kernel regularization on Assist2009. The setting (0.01,0.1) is selected.
Personalized exercise recommendation dynamically aligns pedagogical resources with individual knowledge mastery, which is crucial for satisfying students' dynamic learning needs in modern education. The field is currently driven by two dominant paradigms: Item-Level Exercise Recommendation (ILER) optimizes for immediate single-step state transitions, while Path-Level Exercise Recommendation (PLER) constructs coherent learning paths to maximize cumulative gains. Despite sharing the same ultimate objective, disparate evaluation setups have kept these two lines of research isolated, hindering unified benchmarking and fair comparison. To fill the gap, in this paper, we present a Unified Benchmark for Exercise Recommendation (UniER), a comprehensive evaluation framework that unifies ILER and PLER. Specifically, we introduce Weighted Cognitive Gain (WCG) as a unified metric to measure cross-paradigm algorithmic performance. Our benchmark encompasses 9 datasets spanning four generation methods, facilitating the comparison of 18 representative ILER/PLER methods. Through multi-dimensional analyses covering effectiveness, generalizability, robustness, and efficiency, our results reveal the systematic dominance of PLER and expose the pedagogical failure of ILER's fragmented recommendations under extreme sparsity and noise. Furthermore, we provide an open-source codebase of UniER to foster reproducible research and outline potential directions for future investigations.
Xinghe Cheng, Guiyong Zhuang, Yusheng Xie +5
Jinan University, Guangzhou, China · Beijing University of Technology, Beijing, China · Griffith University, Gold Coast, Australia
Hypergraph can capture complex and higher-order dependencies among learners and learning resources in personalized educational recommender systems. Many existing hypergraph-based recommendation approaches underexplored the dynamic behavioral processes inherent to learning and often oversimplified the complementary information embedded across multiple dimensions (i.e. views) within hypergraphs. These limitations compromise both the distinctiveness of learned representations and the model's generalization capabilities, especially under data-sparse conditions typical in educational settings. In this study, we propose a unified model comprising a dynamic behavioral profiling module and a multi-view attention fusion module based on heterogeneous hypergraph construction. The dynamic behavioral profiling module is designed to capture evolving behavioral processes and infer latent higher-order relations crucial for hypergraph completion; The multi-view fusion module cohesively integrates information from distinct relational views, enriching the overall data representation. The proposed model was systematically evaluated on five public benchmark datasets and one real-world, self-constructed dataset. Experimental results demonstrate that the model outperforms baseline methods across most datasets in key metrics; Furthermore, hypergraph completion based on dynamic behavioral profiling contributes significantly to performance gains, though its efficacy is modulated by dataset characteristics. Beyond offline experiments, we implemented a functional prototype system tailored for postgraduate student literature recommendation. A mixed-methods user study was conducted to assess its practical utility. Quantitative analysis revealed significantly higher perceived recommendation quality; Qualitative feedback highlighted enhanced user engagement and satisfaction with the prototype system.
Tao Xie, Yan Li, Yongpan Sheng +1
Tao Xie, Yan Li, and Jian Liao are with the Faculty of Education, Southwest University, Chongqing, China 400715 · Yongpan Sheng is with the College of Computer and Information Science, Southwest University, Chongqing, China 400715
Learning Path Recommendation (LPR) is critical for personalized education, yet current methods often fail to account for historical interaction uncertainty (e.g., lucky guesses or accidental slips) and lack adaptability to diverse learning goals. We propose U-GLAD (Uncertainty-aware Generative Learning Path Recommendation with Cognition-Adaptive Diffusion). To address representation bias, the framework models cognitive states as probability distributions, capturing the learner's underlying true state via a Gaussian LSTM. To ensure highly personalized recommendation, a goal-oriented concept encoder utilizes multi-head attention and objective-specific transformations to dynamically align concept semantics with individual learning goals, generating uniquely tailored embeddings. Unlike traditional discriminative ranking approaches, our model employs a generative diffusion model to predict the latent representation of the next optimal concept. Extensive evaluations on three public datasets demonstrate that U-GLAD significantly outperforms representative baselines. Further analyses confirm its superior capability in perceiving interaction uncertainty and providing stable, goal-driven recommendation paths.