A newly developed hybrid collaborative filtering (HCF) model significantly improves the accuracy and relevance of physical education (PE) course recommendations by synthesizing multidimensional student digital portraits with data-driven interaction patterns. By integrating basic demographic data with granular information on physical fitness, learning behaviors, and interest preferences, researchers have demonstrated that this dual-branch recommendation framework can successfully mitigate common teaching challenges such as blind course selection, resource mismatch, and low engagement levels. This study, published in November 2026, provides a technical foundation for transforming PE instruction from a one-size-fits-all model toward personalized, data-supported guidance. The experimental results, validated across multiple large-scale datasets, suggest that matching students to courses based on their specific physical capacities and engagement patterns improves overall learning outcomes and student satisfaction.
Traditional PE teaching models frequently suffer from a reliance on subjective course popularity rather than individual student needs. Students often struggle to identify courses that align with their specific fitness levels or sports interests, leading to poor selection outcomes and diminished motivation. Teachers, meanwhile, often lack the dynamic data support necessary to provide targeted feedback or manage teaching resources efficiently. Unlike general academic recommendation scenarios, PE instruction involves a multi-constrained and multi-dimensionally coupled environment where course participation is influenced not just by interest but by the compatibility of a student’s physical fitness, exercise load tolerance, and current performance levels. Existing recommendation frameworks in e-commerce or online learning have historically struggled to characterize these PE-specific constraints adequately, often relying on single-rating metrics that fail to account for the sparsity and uneven distribution of course interaction data.
The HCF model addresses these complexities by constructing a comprehensive digital portrait for every student, which serves as the foundation for the recommendation process. The portrait structure incorporates four primary components: basic attributes, physical ability profiles, learning behavior records, and stated interest preferences. Physical ability data—such as Body Mass Index (BMI), sprint times, long jump distances, and vital capacity—are assigned higher discriminative weights compared to general demographic information to ensure the recommendation engine prioritizes physical compatibility. Furthermore, the model incorporates implicit feedback mechanisms, such as course participation duration, to infer potential preferences objectively. By employing an adaptive feature weight fusion method, the system dynamically adjusts the contribution of these various dimensions based on the data richness of each user, ensuring that even students with limited interaction history receive relevant recommendations.
To refine the user similarity modeling, the researchers introduced a confidence-aware mechanism that assesses the consistency of interest between users during their course interactions. Rather than simply relying on historical course selection—which may overlap due to general popularity—the model implements an evaluation consistency constraint. This mechanism filters common interactive courses by comparing user evaluation scores; if the difference between two users’ ratings for a shared course exceeds a predefined threshold, the course is excluded from the similarity calculation. This ensures that the constructed user similarity relationships reflect authentic, shared interest patterns rather than coincidental enrollment in common courses. The combination of digital portrait-level feature similarity and confidence-aware interest consistency provides a stable and robust basis for predicting a target user’s preference for future PE courses.
The recommendation framework further integrates a course-side similarity component designed to correct for the inherent bias toward popular courses. In many educational environments, basic fitness training or introductory sports courses garner disproportionately high enrollment, which often skews recommendation outcomes. The HCF model utilizes a popularity correction factor, which employs a logarithmic attenuation function for courses that exceed a specific popularity threshold. By suppressing the dominance of high-enrollment courses while maintaining the contribution of specialized, less-populated courses, the model effectively balances course resource distribution. This dual-branch architecture, which fuses user-side feature modeling with course-side popularity correction, ensures that the resulting recommendation list is both technically sound and practically aligned with diverse student cohorts.
The effectiveness of the HCF model was validated through rigorous testing on three distinct datasets: the public Teddy Cup educational platform dataset, a self-constructed PE course dataset gathered from 856 students over a 16-week semester at Qinhuangdao Vocational and Technical College, and the external MovieLens-1M dataset for generalizability. During the preprocessing phase, missing records in the self-constructed PE dataset were addressed using mean imputation and historical behavior-based supplementation, while continuous physical indicators were normalized. These experimental datasets allowed the researchers to evaluate the model’s performance in diverse environments, ranging from general course recommendation to highly specific, context-aware physical education scenarios.
Performance was measured using several key metrics, including Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Accuracy@10, F1-score@10, and Normalized Discounted Cumulative Gain (NDCG)@10. Across all tests, the HCF model demonstrated consistent superiority over established baselines, including User-based Collaborative Filtering, Content-Based Recommendation, Light Graph Convolution Network (LightGCN), Fair Graph Collaborative Filtering, and dual-level graph contrastive collaborative filtering (DL-GCL). On the self-constructed PE dataset, the HCF model achieved an MAE of 0.6763 and an NDCG@10 of 0.8265, representing a 1.94% reduction in MAE and a 2.22% improvement in NDCG@10 compared to the previously best-performing model, DL-GCL. Statistical validation using 5-fold cross-validation and paired t-tests confirmed that these improvements were significant, with p-values consistently below 0.05 and medium-to-large Cohen’s d effect sizes.
A specific ablation study was conducted to isolate the impact of individual sub-modules within the HCF model. By systematically removing the digital portrait module, the confidence mechanism, and the popularity correction factor, the researchers confirmed the synergistic effect of the full model. The results indicated that the digital portrait module provided the most significant foundation for personalization, while the confidence mechanism played a critical role in neighborhood screening. Although the performance degradation after removing popularity correction was relatively smaller than other components, the drop still confirmed that popular course bias exerts measurable interference. These results underscore that the model’s performance is not the result of a single optimization but stems from the integrated, multi-source modeling of user characteristics and course resource distributions.
The study also assessed the model’s performance in cold-start scenarios, defined by users with fewer than five historical course interactions. Given that cold-start users often suffer from the highest recommendation error rates, the HCF model’s reliance on digital portrait attributes proved particularly beneficial. Compared to baseline models, the HCF model achieved a 4.76% reduction in MAE and a 5.05% improvement in NDCG@10 in the cold-start user subset, highlighting the ability of the digital portrait to bridge the data sparsity gap. This confirms that the model’s design—specifically the integration of static physical and demographic information—enables more accurate initial recommendations for new students compared to models that rely purely on historical behavioral interaction graphs.
To evaluate real-world utility, a teaching validation experiment was conducted with 120 students over 12 weeks. An experimental group received personalized course recommendations generated by the HCF model, while a control group relied on traditional independent course selection. The results were striking: the experimental group achieved a course completion rate of 92.63%, significantly higher than the 81.47% observed in the control group. Furthermore, average course participation duration increased by 22.99% in the experimental group. Students in the experimental group reported higher satisfaction levels, noting that personalized recommendations reduced the ambiguity of course selection and allowed them to align their interests with their physical capabilities more effectively.
Despite these successes, the researchers identified current limitations regarding the scope of the captured data. The model currently processes structured data like grades and attendance but lacks real-time, fine-grained dynamic feedback from classroom activity or physiological monitoring. Features such as heart rate or real-time exercise load remain excluded from the current framework, which may limit the system’s ability to adapt to momentary changes in a student’s health status during a specific session. Additionally, the weights for some features are still determined by manual rule sets rather than fully automated representation learning. Future research aims to incorporate deep neural networks and attention mechanisms to reduce reliance on manual rules and broaden data collection to include continuous, long-term learning trajectories.
Looking forward, the authors emphasize the need for large-scale, cross-institutional validation. Since curriculum systems, student demographics, and teaching models can vary substantially between institutions, the stability of the HCF model across different data distributions remains a target for subsequent work. Future trials will focus on deploying the model in various educational contexts to further evaluate its generalization capability and operational stability. By moving toward automated feature learning and integrating dynamic physiological signals, the next generation of HCF systems may provide an even more responsive and personalized approach to physical education, effectively merging the latest technological advancements in recommendation science with established educational theories regarding autonomy and perceived competence.
Subject of Research: Intelligent recommendation systems for personalized physical education course selection.
Article Title: Digital portrait and intelligent recommendation system for physical education teaching based on hybrid collaborative filtering
Article References: Qi, P., Xue, Y., & Hou, Z. (2026). Digital portrait and intelligent recommendation system for physical education teaching based on hybrid collaborative filtering. Discover Artificial Intelligence, 6(1), Article 1895. https://doi.org/10.1007/s44163-026-02340-7
Image Credits: AI Generated
DOI: 10.1007/s44163-026-02340-7
Keywords: Physical Education, Collaborative Filtering, Digital Portraits, Course Recommendation, Student Engagement, Machine Learning, Educational Technology, Personalized Learning, Hybrid Recommendation, User Modeling, Physical Fitness, Recommendation Algorithms, Academic Performance, Teaching Innovation, Learning Analytics, Data Sparsity, Cold-Start Problem, Self-Determination Theory, Human-Computer Interaction
News Source: Courtney Benton. (November 18, 2026). Hybrid Collaborative Filtering Enhances Personalized Physical Education Course Recommendations. Scienmag.



