Reliable Prerequisite-Aware Learning Path Recommendation under Uncertain Knowledge Structures: Long-Horizon Evaluation and Cross-Dataset Validation
Organizations: Department of Computer Science, American International University-Bangladesh (AIUB), Dhaka, Bangladesh · Department of Computer Science and Engineering, Mälardalen University, Västerås, Västmanland, Sweden
Abstract
Learning-path recommendation often depends on knowledge and prerequisite structures whose reliability is uncertain, while high accuracy in predicting the next resource does not necessarily establish reliable long-horizon path behaviour. This study develops a reliability-oriented prerequisite-aware framework that jointly examines immediate next-resource ranking, uncertain prerequisite structures, hard, soft and hybrid constraint semantics, graph robustness, and autoregressive path generation under course-level confirmatory evaluation and independent cross-dataset validation. Development used 3,187 MOOCCubeX courses, followed by a frozen confirmatory evaluation on 541 held-out courses and 37,747 transitions. The primary MOOCCubeX task is structural and course-prefix-conditioned rather than learner-personalised. Junyi provides independent structure-/hierarchy-aware external validation, uses hierarchy metadata rather than an explicit prerequisite graph and is not a zero-shot transfer setting. On the frozen MOOCCubeX evaluation, XGBoost pairwise was the strongest observed immediate ranker, with mean reciprocal rank (MRR) of 0.9048, whereas the proposed prerequisite-aware graph model achieved 0.7707. In the controlled graph-variant development ablation, the no-graph condition produced the strongest immediate ranking result, and immediate prediction showed strong dependence on permitted title/path-similarity information. Autoregressive evaluation exposed path properties not captured by one-step ranking. On Junyi, Markov achieved the strongest observed MRR of 0.4453, compared with 0.3954 for the proposed graph model, with metric-specific differences across ranking measures. Overall, immediate predictive accuracy, prerequisite-structure reliability, and long-horizon path quality should be treated as complementary rather than interchangeable evaluation dimensions in educational recommendation.
Figures & tables
| Method / year | Primary task | Learner state | Knowledge structure | Recommendation / path mechanism | Evaluation focus | Relevance / distinction |
| Deng et al. (2023) | Online course recommendation | Learning-sequence representation | Course knowledge graph with heterogeneous course information | TransD embeddings with BiLSTM, CNN and MLP | HR, MRR and NDCG for course recommendation on MOOCCube | Knowledge-aware sequential ranking; not complete-path evaluation |
| Yun et al. (2024) | Learning-path recommendation | Historical learner records and educational state representation | Prerequisite relations; a prerequisite graph is used in the educational data setting | Doubly constrained offline deep Q-learning | Offline RL evaluation on D4RL and educational datasets | Constrained sequential decision making; emphasis on extrapolation error |
| Ma et al. (2024) | Learning-path recommendation | Knowledge background, learning style, settings and preferences | Knowledge-graph-based multi-behaviour representation | Multi-behaviour transformer with cascading DQN and two-level reward | Experiments on two real online-learning datasets | Learner-adaptive long-/short-term objective modelling |
| Fan et al. (2026) | Learning-path recommendation | Cognitive state and forgetting information | Knowledge graph with antecedent–successor relationships among concepts | MemGNN concept prediction followed by reinforcement-learning sequencing | Three-dataset evaluation and case analysis | Explicitly combines learner-state dynamics, graph structure and path sequencing |
| Zeng et al. (2027) | Concept- and exercise-level learning-path recommendation | Evolving proficiency represented by a bidirectional Mamba-based sequence encoder | LLM-assisted prerequisite and semantic-similarity graphs with directed structural modelling | Hierarchical concept planning with dual-pool adaptive exercise selection | Simulator-based offline evaluation on three real-world datasets | Directly separates behavioural/statistical relations from pedagogical structural validity |
| Dataset / split | Study role | Sequence unit | Split and evaluation design | Candidate construction | Knowledge structure | Sequential protocol |
| MOOCCubeX development Yu et al. (2021) | Model and configuration development | Authored course-resource sequence | 3,187 courses; repeated course-level cross-validation; final development-only train/tune fitting | Full global catalog; training pool of 256; validation cap of 512; development-evaluation cap of 1,024 | Original, validated, refined, and no-graph prerequisite variants | Development-only rollouts and controlled ablations |
| MOOCCubeX frozen confirmatory set | Frozen confirmatory evaluation | Authored course-resource sequence | 541 disjoint courses; all 37,747 transitions evaluated; no transition sampling | Full global catalog; cap of 1,024; development and confirmatory-set catalogs | Frozen refined prerequisite graph with prespecified perturbations | 60 deterministically selected courses; beam search enabled for the first 20; maximum 50 steps |
| Junyi Academy Pojen et al. (2020) | Independent external validation | Learner exercise sequence | 36,269 training, 5,255 validation, and 10,151 test learners; 21,083 short histories excluded | Full 1,330-item catalog at test time; current exercise excluded | Info_Content hierarchy/tree metadata; not treated as a prerequisite graph | 1,000 eligible test learners; rollout horizon of 10 |
| Model | Model family | Graph usage | Features / inputs | Constraint handling | Comparison purpose |
| Random-valid | Heuristic | None | None | Candidate feasibility filter | Random feasible baseline |
| Concept continuity | Structural/content heuristic | Prerequisite depth and other static graph-derived quantities | Selected leakage-safe features | Candidate feasibility filter | Local concept-continuity baseline |
| Popularity | Statistical/content heuristic | None | Training-set popularity, continuity, and gain features | Candidate feasibility filter | Frequency- and content-based baseline |
| Linear HACR-safe | Linear pairwise ranker | Leakage-safe graph-derived features | Full leakage-safe feature set | Candidate feasibility filter | Linear leakage-controlled comparator |
| XGBoost pairwise | Tree-based learning-to-rank | Leakage-safe graph-derived features | Full leakage-safe feature set | Candidate feasibility filter | Nonlinear pairwise comparator |
| Two-Tower | Neural semantic ranker | No graph propagation | Semantic representations and leakage-safe features | Candidate feasibility filter | Neural non-graph comparator |
| Model | HR@1 | HR@5 | HR@10 | MRR | NDCG@10 | Viol@1 | Course-cluster MRR [95% CI] |
| Random-valid | 0.0012 | 0.0048 | 0.0099 | 0.0076 | 0.0047 | 0.0000 | 0.0072 [0.0068, 0.0077] |
| Popularity-content | 0.0001 | 0.0004 | 0.0007 | 0.0044 | 0.0004 | 0.0000 | 0.0039 [0.0035, 0.0044] |
| Concept-continuity | 0.0456 | 0.0565 | 0.0588 | 0.0530 | 0.0523 | 0.0000 | 0.0442 [0.0353, 0.0538] |
| Linear HACR-safe | 0.8069 | 0.8448 | 0.8559 | 0.8259 | 0.8318 | 0.0000 | 0.7985 [0.7831, 0.8140] |
| Two-Tower | 0.7996 | 0.8498 | 0.8616 | 0.8243 | 0.8321 | 0.0000 | 0.8023 [0.7864, 0.8178] |
| Proposed graph | 0.7269 | 0.8207 | 0.8371 | 0.7707 | 0.7853 | 0.0000 | 0.7487 [0.7321, 0.7647] |
| Model treatment | Condition | MRR | Infeasible-target rate |
| A. Prerequisite-graph variants — equal-budget development ablation | |||
| – | No graph | 0.8552 | 0.0000 |
| – | Original | 0.8462 | 0.0096 |
| – | Validated | 0.8337 | 0.0046 |
| – | Refined | 0.8230 | 0.0046 |
| B. Constraint policies — frozen inference and equal-budget retraining | |||
| Model | Rollout | Realised courses | Goal coverage | Termination success | Prereq. violation | Redundancy | Concept diversity / step | Path efficiency |
| Concept-continuity | Greedy | 12 | 0.5999 | 0.0833 | 0.0000 | 0.0596 | 20.2900 | 2.2367 |
| Random-valid | Greedy | 12 | 0.5254 | 0.0833 | 0.0000 | 0.0157 | 5.6750 | 1.5617 |
| Proposed graph | Greedy | 12 | 0.2838 | 0.0833 | 0.0000 | 0.0843 | 1.8933 | 0.5583 |
| Proposed graph | Beam | 4 | 0.4004 | 0.0000 | 0.0000 | 0.1240 | 2.9200 | 2.1850 |
| Panel A: full-catalog ranking | ||||||
| Model | HR@1 | HR@5 | HR@10 | MRR | NDCG@10 | MRR 95% CI |
| Random | 0.0007 | 0.0035 | 0.0075 | 0.0057 | 0.0033 | [0.0056, 0.0059] |
| Popularity | 0.0262 | 0.0860 | 0.1294 | 0.0610 | 0.0706 | [0.0587, 0.0635] |
| Hierarchy continuity | 0.0197 | 0.0981 | 0.1852 | 0.0760 | 0.0857 | [0.0743, 0.0776] |
| Two-Tower no-graph | 0.2158 | 0.6336 | 0.7803 | 0.3946 | 0.4797 | [0.3912, 0.3980] |
| Proposed graph | 0.2170 | 0.6355 | 0.7802 | 0.3954 | 0.4802 | [0.3923, 0.3988] |
Recommender Systems
Recommender systems predict and rank items for people using interaction histories, content, context, and feedback.
See all 573 papers