MODERN APPROACHES TO PERSONALIZED LEARNING PATH PLANNING: KNOWLEDGE GRAPHS, COGNITIVE DIAGNOSIS, NEURAL NETWORKS, AND LARGE LANGUAGE MODELS

Abstract

This study presents a systematic critical review of contemporary approaches to personalized learning path planning in digital education, with particular attention to methods that account for individual differences among students, their current knowledge levels, and their educational goals. The need for personalization is driven by the growing scale of online learning, the high variability in learners’ preparedness, and the diversity of educational content formats. Six key groups of approaches are analyzed, reflecting the evolution of research in this domain. Knowledge graph models, particularly context-aware KGs, provide a formal representation of subject areas, prerequisites, and semantic relationships, thereby forming the basis for constructing cognitively coherent learning paths. Cognitive diagnostic methods (DINA, V-DINA) enable the assessment of mastery of specific concepts and the identification of knowledge gaps, although they require high-quality and comprehensive interaction logs. Combinatorial-optimization schemes, such as Fuzzy-CDF combined with the Apriori algorithm and particle swarm optimization (PSO), implement the sequencing of learning concepts and the selection of resources, ensuring a balance between prerequisites, cognitive capabilities, and the learner’s preferred style. Another strand is represented by fuzzy-neural models (fuzzy-ANN), which build on Kolb’s Learning Style Inventory (LSI) and integrate fuzzy weights with the adaptability of artificial neural networks, thus allowing for the consideration of individual educational strategies. Sequential and multitask architectures (Seq2Seq, LSTM with non-repeat loss) focus on predicting the learner’s next steps while simultaneously implementing knowledge tracing, which makes it possible to monitor the probability of successful mastery of material. A promising direction is the use of multimodal approaches that integrate various types of data (text, video, behavioral signals), along with large language models (LLMs), which provide semantic enrichment of knowledge graphs, interpretation of natural language queries, and explainability of recommendations.

The analysis highlights the strengths and weaknesses of each approach: from the high degree of structure and explainability of graph-based methods to the data and resource demands of neural networks; from the accuracy of cognitive diagnostics to its dependence on log completeness; from the flexibility of LLMs to the challenges posed by hallucination risks and the need for quality control. The review shows that none of the approaches is universal, and the most promising direction lies in building hybrid architectures that combine the structured nature of knowledge graphs, the diagnostic capacity of cognitive methods, the adaptability of neural networks, and the flexibility of LLMs. Future research should focus on the integration of multimodal signals, the development of reproducibility protocols, the unification of evaluation metrics, and the construction of scalable systems capable of operating effectively in large-scale educational environments such as MOOCs.