💻 computer science

Construction of a Personalized Physical Fitness Training Recommendation System Based on Random Forest and Gradient Boosting Machine

This paper proposes a Domain-Knowledge-Driven Hierarchical Collaborative Learning (DKHCL) framework that integrates Random Forest feature selection with Gradient Boosting Machine prediction and Acute-Chronic Workload Ratio constraints to achieve high-accuracy personalized physical fitness recommendations, demonstrating both superior offline performance against advanced baselines and significant real-world improvements in injury reduction and goal-achievement efficiency in a prospective field validation.

Original authors: Fuliang Geng, Mei Wei

Published 2026-08-26

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Fuliang Geng, Mei Wei

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Every athlete, from the weekend jogger to the Olympic sprinter, faces a fundamental biological puzzle: how much stress to apply to the body to get stronger without breaking it. Push too little, and the body adapts slowly; push too hard, and the risk of injury skyrockets. For decades, coaches have relied on intuition and experience to solve this, trying to balance an athlete's current fatigue against their long-term fitness. However, human judgment often struggles to account for the complex, shifting history of an athlete's training load. A new approach, developed by researchers at Laiwu Vocational Technical College in China, uses artificial intelligence to navigate this delicate balance, transforming how personalized fitness plans are created. The core idea relies on two established concepts in sports science. First, the body's response to exercise is not just about the workout happening today, but how that workout compares to what the athlete has done over the last few weeks. Second, the relationship between training and recovery is not a straight line; it involves non-linear patterns where small changes in intensity can lead to disproportionately large effects on performance or injury risk. By teaching a computer to recognize these patterns while strictly adhering to safety rules derived from human physiology, the researchers have built a system that suggests training programs with a level of precision previously difficult to achieve.

The researchers set out to build a system that could look at a person's detailed physical profile and recommend the perfect training program from a library of options. They gathered data from 2,634 individuals over an 18-month period, drawing from a provincial sports center, three fitness clubs, and a rehabilitation facility. For each person, they collected 37 different pieces of information, ranging from static details like age and height to dynamic physiological markers such as heart rate variability, muscle strength, and oxygen uptake. They also tracked how these individuals trained over time, recording the frequency, duration, and intensity of their workouts. The goal was not just to predict a score, but to determine the "suitability" of a specific training plan for a specific person on a five-point scale, from highly unsuitable to highly suitable. To create the ground truth for this task, a panel of five expert sports scientists and coaches independently rated every possible combination of athlete and program. When the experts disagreed, they discussed the cases until they reached a consensus, ensuring the data the computer learned from was as reliable as human expertise could make it.

To solve this problem, the team constructed a two-part artificial intelligence framework that works differently than standard machine learning models. The first part acts as a filter, using a method called Random Forest to sift through the 37 available data points and identify the 15 most important ones. This step is crucial because it removes the noise and focuses the system on the signals that truly matter, such as an athlete's maximum oxygen capacity and their recent training frequency. The second part, a Gradient Boosting Machine, takes these 15 key features and makes the actual prediction. What makes this system unique is how it handles time and safety. Instead of trying to learn the passage of time from raw data alone, the researchers explicitly fed the system calculated ratios that compare an athlete's recent one-week workload against their four-week workload. This ratio, known as the acute-chronic workload ratio, is a well-known safety metric in sports science. The researchers did not just let the computer learn this on its own; they built a rule directly into the learning process that penalizes the system if it suggests a plan that would push this safety ratio too high. This ensures that the computer never recommends a dangerous increase in training load, even if the data might otherwise suggest it would lead to a high score.

The results of this approach were tested against a wide range of other methods, including single machine learning models, deep learning systems that mimic the human brain, and traditional expert-based planning. In offline tests using a fixed set of data, the new system correctly identified the suitability of training programs 87.3% of the time. This performance was higher than the best deep learning models, which reached 85.8%, and significantly better than traditional methods. The system also proved to be much faster to train, taking only about 52 seconds to process the data on a standard workstation, whereas the deep learning models required over three minutes. More importantly, the researchers took the system out of the lab and into the real world for a 16-week field validation involving 600 participants. It is important to note that this was a prospective non-randomized controlled cohort study, meaning group allocation was based on athlete preference rather than random assignment, which introduces potential selection bias. In this real-world test, the group using the system's recommendations saw their injury incidence rate drop to 3.2 per 1000 exposure hours, a figure much lower than the 7.1 observed in groups following traditional training plans. Furthermore, the athletes using the system reached their fitness goals an average of 21 days faster than the control groups.

The study suggests that the most effective way to apply artificial intelligence in sports training is not to let the computer learn everything from scratch, but to guide it with established human knowledge. By embedding physiological safety rules directly into the algorithm's learning process, the system avoids the common pitfall of suggesting dangerous training loads just because the data looks promising. The researchers found that while deep learning models are powerful, they require massive amounts of data to learn these safety constraints on their own, whereas their hybrid system achieved superior results with a much smaller dataset by starting with the right rules. The system also incorporates a dynamic adjustment mechanism: if an athlete's feedback indicates high fatigue or poor recovery over a specific period, the system triggers a re-evaluation that can reduce the recommended training intensity by 10–15%. This creates a closed loop where the training plan evolves with the athlete's daily condition, rather than remaining a static schedule. While the study was conducted with a specific group of participants and focused on short-term outcomes, the findings offer a compelling blueprint for how technology can support, rather than replace, the nuanced judgment of human coaches, making high-level training safer and more efficient for a broader range of people.

Technical Summary: Construction of a Personalized Physical Fitness Training Recommendation System Based on Random Forest and Gradient Boosting Machine

1. Problem Statement

Traditional physical fitness training programs often rely on coach experience, struggling to balance individual physiological differences with dynamic adjustment needs. Existing machine learning approaches for personalized recommendations face three fundamental limitations:

Neglect of Temporal Dependencies: Most systems treat training load as a static optimization problem, failing to account for how recent training history influences current fatigue and adaptation. While Acute-Chronic Workload Ratio (ACWR) theory exists, it is typically applied as a post-hoc filter rather than being embedded within the learning process, leading to predictions that may violate physiological safety domains.
Accuracy-Robustness Trade-off: Single algorithms struggle to balance performance and stability. Random Forest (RF) offers noise tolerance but lacks fine-grained residual characterization, while Gradient Boosting Machines (GBM) achieve high precision but are sensitive to feature noise.
Ineffective Fusion Strategies: Existing ensemble methods often rely on simple linear combinations or meta-learners (stacking) that fail to leverage the complementary structural properties of different algorithms or incorporate domain-specific confidence measures dynamically.

2. Methodology: The DKHCL Framework

The paper proposes a Domain-Knowledge-Driven Hierarchical Collaborative Learning (DKHCL) framework. This architecture integrates RF for feature selection and GBM for ordinal prediction, constrained by exercise physiology principles.

2.1 Data Pipeline and Preprocessing

Dataset: 2,634 valid samples from 2,847 trainees (collected over 18 months from a provincial sports center, fitness clubs, and a rehabilitation center).
Features: Raw data includes 37 dimensions (static attributes, dynamic physiological indicators, and behavioral tracking).
Temporal Encoding: Instead of native sequential modeling (e.g., RNNs), temporal dependencies are encoded via engineered rolling-window statistics: 7-day acute load, 28-day chronic load, and ACWR ratios.
Labeling: Suitability is defined as an ordinal classification task (5-point Likert scale: 1=Highly Unsuitable to 5=Highly Suitable), labeled by an expert panel (ICC = 0.87).

2.2 Core Components

The framework operates in three distinct stages:

RF Feature Selection (Dimensionality Reduction):
- A Random Forest model (RF-FS) is trained on the full 37-dimensional set to identify the most predictive features using conditional permutation importance.
- This reduces the input space to a core subset of 15 features (e.g., VO₂max, muscle strength index, training frequency), compressing the feature space by ~55% while retaining 89.7% of cumulative predictive information.
- This step acts as a noise filter, preparing a cleaner input space for the predictor.
GBM Ordinal Prediction with ACWR Constraints:
- A Gradient Boosting Machine (GBM) performs the primary ordinal classification on the reduced 15-dimensional feature set.
- Domain Knowledge Injection: Unlike standard GBM, this model embeds the ACWR safety constraint directly into the training process via a gradient penalty.
- If a predicted training program implies an ACWR > 1.5 (a physiological risk threshold), a soft penalty term is added to the loss function during gradient iteration. This ensures the model learns to avoid unsafe regions inherently, rather than clipping predictions post-hoc.
- Hyperparameters are optimized using Bayesian optimization (Optuna).
Dynamic Hierarchical Fusion:
- The final recommendation score is a weighted combination of RF and GBM outputs: $Score = w_{RF} \cdot y_{RF} + w_{GBM} \cdot y_{GBM} + \sum \theta_k \cdot g_k$ .
- Uncertainty-Adaptive Weights: The weights ( $w_{RF}, w_{GBM}$ ) are not fixed. They are dynamically calculated based on the prediction variance (uncertainty) of each model for a specific sample. If RF is more confident (lower variance) for a specific input, its weight increases, and vice versa.
- Rule-Based Intervention: A strategy layer applies hard constraints (e.g., medical contraindications, equipment availability) and diversity rules before final ranking.

3. Key Contributions

Physiologically Constrained Learning: The paper introduces a mechanism to embed ACWR theory as a soft penalty term within the GBM gradient boosting process, ensuring predictions adhere to physiological safety boundaries during training rather than after.
Hierarchical Collaborative Architecture: It proposes a specific "RF for feature selection $\to$ GBM for refinement" pipeline. This decouples dimensionality reduction from prediction, leveraging RF's noise robustness to stabilize the GBM's gradient descent.
Uncertainty-Driven Dynamic Fusion: The system moves beyond fixed-weight ensembles by calculating sample-specific fusion weights based on model variance, allowing the system to adaptively trust the most reliable model for each specific trainee profile.
Temporal Feature Engineering: The study demonstrates that explicit engineering of temporal features (rolling windows, ACWR) can outperform end-to-end deep sequential models (LSTM/Transformer) on tabular data with limited sample sizes, offering better interpretability and computational efficiency.

4. Experimental Results

4.1 Offline Performance

Experiments were conducted using 5×2 cross-validation on a fixed train-validation-test split.

Accuracy: DKHCL achieved 87.3% (95% CI: 85.1–89.4%), outperforming single models (RF: 74.7%, GBM: 81.5%) and deep learning baselines (Transformer: 85.8%).
AUC: DKHCL achieved 0.924, compared to 0.918 for Transformer.
F1-Score: DKHCL achieved 0.883.
Efficiency: DKHCL training time was 52.4 seconds per fold, significantly faster than Transformer (203.4s) and LSTM (156.2s).
Statistical Significance: While DKHCL showed a 1.5% accuracy improvement over the strongest baseline (Transformer), the paper notes this difference did not reach statistical significance after Bonferroni correction for multiple comparisons, though the effect size (Cohen's d = 0.18) suggests a practical advantage.

4.2 Field Validation (Prospective Cohort)

A separate 16-week prospective study with 600 participants (distinct from the offline dataset) evaluated real-world outcomes, comprising an Experimental group (Hybrid System, n=180), Control Group A (Traditional, n=90), and Control Group B (Single RF, n=90):

Injury Rate: The system-guided group had an injury rate of 3.2%, compared to 7.1% in traditional control groups.
Goal Achievement: The mean goal achievement cycle was shortened by 21 days relative to controls.
Adoption: Plan adoption rates reached 89.4%, significantly higher than traditional single-dimensional profiles (65.7%).

5. Significance and Claims

The paper claims that the DKHCL framework demonstrates that explicitly embedding domain knowledge (physiological constraints) as inductive biases is more sample-efficient than end-to-end black-box learning in small-sample, high-noise domains like sports science.

Interpretability vs. Performance: The study argues that by engineering temporal features and embedding safety constraints, the system achieves high accuracy and safety without the computational cost and "black-box" nature of deep learning models.
Practical Impact: The system successfully transitions from static planning to dynamic evolution, reducing injury risks and accelerating fitness goals in real-world settings.
Limitations Acknowledged: The authors modestly note that the field validation was a non-randomized controlled cohort study, introducing potential selection bias. They also acknowledge that the system currently relies on specific data sources (wearables, logs) and does not yet incorporate multimodal data like video analysis or psychological states. The ACWR threshold (1.5) is acknowledged as a heuristic derived from team sports, requiring further calibration for specific populations.

In conclusion, the paper presents a robust, interpretable, and computationally efficient solution for personalized fitness training that prioritizes physiological safety and leverages the complementary strengths of tree-based ensemble methods.