Recommender Systems and Learning Paths

Recommender systems and learning paths — the part of adaptive and Personalized Learning technology that decides what a learner should encounter next and in what order: which resource, practice item, or course to rank toward them, and which sequence of concepts to walk through. Its two method families are behavioral — collaborative filtering over interaction logs — and semantic — sequencing over a knowledge graph of concepts, resources, and prerequisite relations — increasingly fused into hybrid models. Because the output is a ranked list rather than a dialogue, the distinctive problems are selectivity and legitimacy: cold start and popularity bias under sparse data, the directional asymmetry of prerequisites, and whether a teacher or learner can understand, audit, and trust the list they are shown.

Questions to Consider

• A recommender learns from what learners clicked, watched, and completed. If the system only sees behavior, whose learning is it actually modeling — and what does it miss for a student who is quiet, struggling, or already knows the material?
• In one study, teachers found curricular-language explanations more understandable and trustworthy than feature-importance charts. What would an explanation of a recommendation have to say before you would act on it?
• Prerequisite relations run in one direction only — knowing B does not imply knowing A. Why might a model that scores concept pairs independently get this wrong, and what would go wrong downstream for a learner if it did?
• Path optimization studies report shorter paths and better post-test scores, with reduced cognitive load as the main mediator. Is a shorter path always the better path, or can efficiency remove practice a learner needed?
• An audit of an educational recommender found popular materials still dominated lists after fairness and diversity interventions. If popularity keeps winning, what would make you redesign the objective rather than tune the reranker?
• Learning Design data from 554 courses showed Acquisition (knowledge transmission) as both the most common activity type and the most common entry point. If AI recommends "what comes next" from that data, whose design habits does it reproduce?

Introduction

Educational recommender systems exist because the supply of learning material has outgrown any human's ability to navigate it. Introductory programming alone has many thousands of practice activities, and organizing them into instructionally meaningful bundles normally requires time-intensive expert curation. A recommender takes a learner's history, the properties of the material, or both, and produces a ranked shortlist or an ordered path through content.

Where Adaptive Learning describes the mechanism of adjusting content, pacing, and difficulty to a learner model, and Intelligent Tutoring describes the platform that does so in dialogue, recommender research concentrates on the selection and sequence decision itself: the ranking function, the structure it ranks over, and the evidence it uses. That decision layer brings its own failure modes, taken up in the sections below.

How Educational Recommenders Work

Behavioral recommenders inherit the logic of collaborative filtering: learners with similar histories are assumed to want similar resources, so the system predicts from co-occurrence rather than content. This is powerful where data are dense and brittle where they are not. Evangelista and Bukhari (2026) state the problem plainly — accuracy-focused recommenders give less reliable support to students with limited participation histories while popular resources dominate lists — and answer it with a Hybrid HKG-GRU framework that embeds a heterogeneous Knowledge Graph of course materials and models learner sequences with a GRU, then trains for robustness across learner groups (GroupDRO) and reranks for exposure diversity (Maximal Marginal Relevance). On Moodle logs from 152 students, 59 resources, and roughly 150,000 interactions it reached HR@10 = 0.68 and MRR = 0.41, but substantial popularity bias persisted at catalog level — evidence that bias mitigation is partial.

Semantic recommenders replace co-occurrence with structure. Liu, Sun, and Song (2026) build a knowledge graph of teaching resources, language concepts, skills, learner groups, and pedagogical attributes, and decompose every resource along four instructional dimensions — teaching context, cognitive level, technological feature, and cultural adaptability. Recommendation expands k-hop over the graph from a learner's stated requirements, refines candidates with a feature-based collaborative filter, and ranks them by a fusion coefficient that rises with the learner's ability, progress, and interest indices, so less-advanced learners get more structured guidance from the graph while higher-ability learners lean on their behavioral pattern. On the English subset of the MARS dataset (4,800 users, 14,200 resources, 132,000 interactions, sparsity 0.9981) it reached NDCG 0.625 and HR 0.751, roughly a 7–8% gain over the best neural baseline, and beat graph-aware baselines including RippleNet and KGAT on coverage and cross-domain accuracy. Ablation shows the signals are complementary (CF-only NDCG 0.584, KG-only 0.604), and the authors note these are ranking metrics, not evidence of learning effectiveness.

A third route avoids per-learner interaction data almost entirely. Hoq and colleagues (2026) extract pattern-based knowledge components from each code sample and recommend related practice activities by the similarity of their knowledge-component sets. On an expert-organized corpus of introductory Python materials, the approach aligned with the instructors' own conceptual bundles and beat knowledge-component and embedding baselines on ranking metrics — evidence that instructional alignment can come from the material's semantics where learner histories are thin.

Prerequisites and Path Sequencing

Ranking a next item is easier than ordering a curriculum, because learning dependencies are directional. Cheng and colleagues (2026) argue that treating prerequisite discovery as ordinary link prediction fails on three properties of the relation: it is irreversible, its evidence is often multi-hop (ci → ck → cj) rather than a one-step learner transition, and its relevance is pair-specific, so a concept's representation should change depending on what it is paired with. Their ProPRL model adds an Irreversibility Constraint — an anti-symmetry regularizer penalizing high confidence in both directions of a pair — alongside a pair-conditioned gate weighting resource-aware evidence from a concept-resource hypergraph against behavior-aware evidence from a learning-behavior graph. It ranked first on all nine dataset–metric combinations across MOOC, LectureBank, and University Course datasets, improving on the strongest baseline by 1.96% to 6.11%, and the constraint raised the share of correctly ordered relations from 88.0% to 90.0%.

Once dependencies are modeled, a path becomes an optimization problem. Feng and Huang (2026) combine Bayesian cognitive diagnosis, knowledge space theory, and cognitive load theory: a Bayesian DINA model converged on EdNet data with 91.3% sparsity, and a shortest remediation path algorithm produced personalized paths averaging 3.82 steps — 22.4% more efficient than random paths and 23.6% more efficient than fixed full-coverage paths. In a randomized probability-learning experiment with 120 students, personalized-path learners finished in 57.6 minutes against 73.8 minutes for controls and outperformed controls on the post-test after controlling for pre-test. The study also tests why: personalized paths reduced cognitive load on all six adapted NASA-TLX dimensions, and cognitive load was the primary mediator of the effect on post-test performance (indirect effect 0.28 in the multiple-mediation model). A Hidden Markov Model identified Analytical Thinking as the bottleneck attribute, with the lowest forward transition probability (0.31) and the highest guessing parameter (g = 0.28).

PersonaPath pushes the sequencing question one level up by disputing the unit of the decision. Liu et al. (2026) contrast Exercise-Centric recommendation — infer the next item from interaction logs — with Knowledge-Centric planning, in which a planner reads an explicit learner persona, a mastery state and a stated target unit and selects the next textbook, unit and concept from a curriculum hierarchy (347 textbooks, 1,751 units, 4,092 concepts across 77 subjects, with 411 prerequisite edges verified to 99.5% precision at Cohen's κ = 0.93). The distinguishing case is the one logs cannot see: two learners with identical correctness records but different targets need different routes, and only the goal makes that visible. Their closed-loop evaluation of ten LLMs (1B–30B+) separates the dimensions an aggregate metric conflates — DeepSeek-V3.1 reaches 90.9% on prerequisite/hallucination validity but only 44.3% on adaptivity, for a 29.5% final pass rate in Basic Education and 14.6% in Higher Education, and no model exceeds 44.7% on adaptivity anywhere. Two ablations sharpen the diagnosis: removing the mastery field from the persona costs up to 26.1 points of adaptivity while barely moving validity, and one-shot path generation raises validity by as much as 30.8 points while cutting adaptivity by 28.8 — curriculum-conformant sequencing is a far easier target than learner-conditioned sequencing, which is a caution against reporting path quality without a learner-alignment constraint (PersonaPath: Towards Knowledge-Centric Personalized Learning Path Planning).

Prerequisite structure also works as a diagnostic lens on the curriculum rather than the learner. Medhat and colleagues (2026) classify student questions to an AI teaching assistant against a GPT-4-extracted prerequisite graph, reaching 80.0% accuracy across 43 labels on 1,340 question events from 164 graduate students, and found topic-level question volume correlated with students' self-reported difficulty (Spearman's ρ = 0.491, p = 0.008). The interaction log becomes a map of where the curriculum's ordering is failing learners, at no added assessment burden.

Explaining and Auditing Recommendations

Ranking quality does not settle whether anyone should follow the ranking. Feldman-Maggor, Cukurova, Kent, and Alexandron (2025) adapt Hoff and Bashir's trust-in-automation model to AI recommendations and test it with 41 in-service chemistry teachers using the recommendation tool GrouPer. Explainability raised Trust indirectly, by raising understandability, and the form of the explanation mattered more than its presence: moving teachers from feature-importance ("data-driven") explanations to semantic, curricular-language ("domain-driven") explanations significantly increased understandability (W = 80.5, p = 0.005), learned trust (W = 52, p = 0.002), and acceptance (W = 22.5, p = 0.003), and all seven think-aloud teachers reported the domain-driven explanations as more influential. Two further acceptance drivers appeared, both situational rather than epistemic: pedagogical alignment (reported by 8 of 11) and workload reduction (6 of 11). Several teachers also said explanation alone was insufficient — they wanted classroom experience with the tool before relying on it, a strong claim that trust in a recommender is built over use, not delivered in a legend.

Fairness auditing asks a different question: whose recommendations are worse? The Graph-GRU study is a useful worked example because it names what its audit could not see. Fairness was operationalized through participation-based cohorts because the public Moodle dataset lacked achievement, prior knowledge, learning profiles, and demographic attributes, so the result is an audit of behavior across engagement levels rather than a full assessment of educational equity; a low-activity learner may be struggling, disengaged, or already familiar with the material. Explainability there takes the form of path-based and counterfactual analysis, with median counterfactual stability CR@10 = 1.0 for many learners but persistent catalog-level popularity bias reflected in high Gini exposure metrics. The lesson is that Bias Mitigation must be measured at the level of exposure and learner groups, not inferred from accuracy or from the presence of an explanation module, and that Human-in-the-Loop oversight is meaningful only if the human sees the group-level evidence.

Learning Paths in Practice

Paths are not only computed; they are designed. Divjak, Svetec, and Horvat (2026) turned Learning Analytics on the design process itself, analyzing the sequence of 29,064 teaching and learning activities across 554 courses planned in a free Balanced Design Planning tool. Markov chains and pattern mining surfaced a design grammar: Acquisition-type activities were the most common learning type (above 20%) and the most common entry point, followed by Practice, Discussion, and Assessment (each 15–20%), and the strongest transition was Assessment → Discussion (0.332), ahead of Practice → Practice (0.317). The rule Acquisition → Assessment → Practice → Practice held at confidence 0.74 (lift 1.45), and learning type tracked the intended outcome's cognitive level: Acquisition fell from about 50% of activities at Bloom level 1 to about 20% at level 6. The authors warn that resemblance to flipped, inquiry-based, or project-based designs is not evidence of intent — but a recommender trained on such design data learns these habits, including a bias toward transmission-oriented opening activities.

At the other end of formal schooling, path structure becomes a AI Governance question. Szekely, Gal-Ezer, and Harel (2026) argue that AI-mediated access to knowledge warrants rethinking fixed higher-education curricula and propose "learnity graphs" — structured representations of learning as interconnected units of knowledge, skills, experience, and artifacts — spanning academic, professional, and personal learning. The proposal keeps the university's role in foundational knowledge while shifting emphasis toward Creativity and interdisciplinary integration, and it makes the learner a maintainer of their own graph, which is where it meets Self-Directed Learning and Self-Regulated Learning. Whether self-directed navigation leads somewhere constructive is not guaranteed: Shen and Arunrugstichai (2026) model two pathways from high-school learning climate to university GenAI use in a cross-contextual survey (N = 508), distinguishing constructive autonomous use from compulsive dependence.

One boundary worth stating: what recommenders currently do is closer to consultation than to collaboration. Mutlu Cukurova (2026) reconstructs what has historically been required before an interaction qualifies as collaborative — a negotiated, partly symmetric relationship, shared and negotiable goals, a low and shifting division of labor, and mutual modeling and socially shared AI Regulation in Education — and concludes that most current human-AI interaction is better described as consultation, governance, delegation, or instruction. His five-level taxonomy (transactional, situational, operational, praxical, synergistic) is a useful calibration: suggesting the next resource is an operational arrangement, and calling it a collaborative learning partner inflates the claim.

How This Page Relates to Personalized, Adaptive, and Tutoring Research

The three adjacent concept pages cover adjacent ground and should be read together with this one. Personalized Learning covers the goal — tailoring the whole experience, including goals, preferences, and pace, of which pathway selection is one component. Adaptive Learning covers the mechanism — the measure-model-adapt loop that changes content, pacing, and difficulty in real time. Intelligent Tutoring covers the platform, the canonical system that combines diagnosis with instructional interaction. This page covers the decision layer those pages presuppose: the ranking and sequencing technology itself. It is the home for collaborative-filtering and knowledge-graph recommender architectures, prerequisite-relation discovery and curriculum ordering, the explainability and fairness of ranked outputs, and the recommender-specific failure modes — cold start, popularity bias, over-narrowing, opaque ranking, and unequal recommendation quality across learner groups. A study reporting NDCG or coverage on interaction logs belongs here; one reporting learning gains from a tutoring dialogue belongs on the tutoring or adaptive-learning page.

Risks and Open Questions

The evidence base has a consistent shape: strong offline ranking metrics, weak evidence about learning. Both recommender studies above are evaluated on historical interaction logs and say so, and the CF–KG authors call explicitly for teacher assessments, learner studies, and outcome-based experiments. Popularity bias survived a fairness-aware training objective in one system, so the field cannot yet point to a reranking or regularization fix that reliably changes exposure, and fairness audits remain bounded by what LMS logs contain. Cold start is addressed architecturally, through semantic structure, but not tested on genuinely new users in live deployment. Path efficiency is better evidenced, yet it rests on one study and one algorithm family, and a shorter path is not automatically better if it removes the productive struggle a learner needs. The vocabulary is also unstable: Ikram and colleagues (2026), in a PRISMA review of 31 Scopus-indexed articles (2013–2025), report medium-to-large cognitive effects (g = 0.50–0.70) for AI-enabled adaptive systems that are heavily moderated by implementation quality and study design.

Connected Concepts

• Knowledge Graph — Structured representations of concepts, resources, and prerequisite relations
• Explainable AI — Explanation as a design feature that mediates understandability and trust
• Curriculum Design — Ordering decisions that recommenders model, infer, and reproduce
• Learning Design — The planned sequence of activities that path analyses operate on
• Learning Analytics — The data and methods behind recommendation, and the object of audit
• Cognitive Diagnosis — Diagnosing knowledge states that seed remediation paths
• Knowledge Tracing — Modeling mastery over time to decide what a learner is ready for
• Learner Modeling and Adaptive Instruction — Learner representations that recommendation consumes
• Personalized Learning — The broader goal; this page covers the selection decision
• Adaptive Learning — The real-time adjustment mechanism; this page covers ranking and sequencing
• Intelligent Tutoring — The platform that acts on recommendations in dialogue
• Bias Mitigation — Fairness interventions in ranking and exposure
• Equity — Group-level differences in recommendation quality
• Human-in-the-Loop — Oversight of automated resource navigation
• Self-Directed Learning — Learner-maintained pathways and lifelong navigation
• Lifelong Learning — Graph-structured learning beyond a degree sequence
• Trust Calibration — Matching reliance to recommendation reliability

Connected Articles

• A Hybrid Collaborative Filtering and Knowledge Graph-Based Cross-Domain Recommendation Method for Multimodal Teaching Resources — Hybrid collaborative filtering plus knowledge graph for multimodal teaching resources, with ability- and progress-aware fusion (Liu, Sun & Song 2026)
• Fair and explainable educational recommendations with a hybrid Graph-GRU framework — Heterogeneous knowledge graph + GRU recommender with GroupDRO fairness and counterfactual explainability (Evangelista & Bukhari 2026)
• The Impact of Explainable AI on Teachers' Trust and Acceptance of AI EdTech Recommendations: The Power of Domain-specific Explanations — Domain-driven explanations build more teacher trust than feature-importance charts (Feldman-Maggor et al. 2025)
• ProPRL: Property-Aware Prerequisite Relation Learning in Educational Knowledge Graphs — Prerequisite relation learning with an irreversibility constraint for directional consistency (Cheng et al. 2026)
• Automated Recommendation of Programming Learning Content Using Pattern-based Knowledge Components — Recommending programming practice by pattern-based knowledge-component similarity (Hoq et al. 2026)
• Learning paths and patterns in learning design: Insights from 500+ courses — Markov and pattern-mining analysis of 29,064 designed activities across 554 courses (Divjak, Svetec & Horvat 2026)
• Bayesian cognitive diagnosis optimizes personalized learning paths via mediation of cognitive load and Hidden Markov Model state transitions — Bayesian DINA diagnosis and shortest remediation paths, mediated by cognitive load (Feng & Huang 2026)
• Detecting Knowledge Gaps from Conversational AI Interactions Using Curriculum Prerequisite Graphs — Mapping AI TA questions to a GPT-4-extracted prerequisite graph to find curriculum-level gaps (Medhat et al. 2026)
• Rethinking Higher Education: From Fixed Curricula to Learnity Graphs — Learnity graphs as a lifelong-learning alternative to fixed curricula (Szekely, Gal-Ezer & Harel 2026)
• What do you mean by human-AI collaboration: Prerequisite functions and the affordances needed to achieve it — Five-level taxonomy showing most human-AI interaction is consultation, not collaboration (Mutlu Cukurova 2026)
• How High-School Pressure and Autonomy Support Are Linked to Dual AI Learning Pathways: A Cross-Contextual SEM Analysis — Autonomy support versus pressure predicting constructive or compulsive AI pathways (Shen & Arunrugstichai 2026)
• Artificial Intelligence in Education: A Systematic Review of Personalized Learning Trends and Future Directions — PRISMA review of personalized learning trends, pathways, and recommendation models (Ikram et al. 2026)
• PersonaPath: Towards Knowledge-Centric Personalized Learning Path Planning — PersonaPath: a Knowledge-Centric planning benchmark showing LLMs reach 90.9% validity but only 44.3% adaptivity (Liu et al. 2026)