FITNESS EXERCISE CLASSIFICATION BASED ON THE XGBOOST MACHINE LEARNING MODEL
Authors:
1.
KNITU-KAI
2.
Kazan State Technical University
(Sistem informacionnoy bezopasnosti, student)
student
Russian Federation
student
Russian Federation
Type:
Article
Pages:
from 113
to 117
Status:
Published
Received:
03.03.2026
Accepted:
08.04.2026
Published:
05.05.2026
Language:
Russian
Keywords:
FITNESS EXERCISE CLASSIFICATION, MACHINE LEARNING, XGBOOST MODEL, TIME SERIES ANALYSIS, MODELING, CLASSIFICATION QUALITY METRICS
Abstract:
This article explores the construction and evaluation of the XGBoost machine learning model for classifying fitness exercises. Assessing the accuracy of human fitness exercise performance is a hot topic in machine learning and artificial intelligence. Machine learning algorithms used to address this issue can identify errors in movement technique, evaluate training effectiveness, and provide personalized recommendations for correcting physical activity and improving the biomechanics of human movement during patient rehabilitation. To address this issue, train, and test the model, the "Physical Therapy Exercises Dataset" from the UCI Machine Learning Repository was selected. The dataset consists of 1,378,015 records collected during gym training sessions involving five subjects performing eight types of physical therapy exercises using inertial measurement units and magnetic sensors. Each of the eight exercise classes is represented by 55 files (12.5% of the dataset), ensuring a balanced sample and eliminating model bias. All data were randomly divided into a training (80%) and a test (20%) set. The XGBoost model was built using the PyCharm integrated development environment and Notepad++. Data decimation was used to optimize performance, significantly reducing the time it took to build and evaluate the model. The user can select a decimation coefficient from 1 to 10. With maximum decimation (sample_rate=10), the sampling frequency is reduced from 25 Hz to 2.5 Hz. This optimization provides a tenfold acceleration in loading (from 3-5 minutes to 20-30 seconds) while maintaining classification accuracy at 90-92%. Data batching was also implemented, involving batch processing of 10,000 records per iteration to reduce the load on RAM. The accuracy of the constructed model was also compared with that of other classification methods. A streamlit application was selected for the interactive presentation of the study results. Among all the models built, XGBoost demonstrated the highest accuracy results, indicating its effectiveness. Therefore, the model can be effectively used as a tool for assessing the correctness of fitness exercises, for example, using a mobile or web app.
This article explores the construction and evaluation of the XGBoost machine learning model for classifying fitness exercises. Assessing the accuracy of human fitness exercise performance is a hot topic in machine learning and artificial intelligence. Machine learning algorithms used to address this issue can identify errors in movement technique, evaluate training effectiveness, and provide personalized recommendations for correcting physical activity and improving the biomechanics of human movement during patient rehabilitation. To address this issue, train, and test the model, the "Physical Therapy Exercises Dataset" from the UCI Machine Learning Repository was selected. The dataset consists of 1,378,015 records collected during gym training sessions involving five subjects performing eight types of physical therapy exercises using inertial measurement units and magnetic sensors. Each of the eight exercise classes is represented by 55 files (12.5% of the dataset), ensuring a balanced sample and eliminating model bias. All data were randomly divided into a training (80%) and a test (20%) set. The XGBoost model was built using the PyCharm integrated development environment and Notepad++. Data decimation was used to optimize performance, significantly reducing the time it took to build and evaluate the model. The user can select a decimation coefficient from 1 to 10. With maximum decimation (sample_rate=10), the sampling frequency is reduced from 25 Hz to 2.5 Hz. This optimization provides a tenfold acceleration in loading (from 3-5 minutes to 20-30 seconds) while maintaining classification accuracy at 90-92%. Data batching was also implemented, involving batch processing of 10,000 records per iteration to reduce the load on RAM. The accuracy of the constructed model was also compared with that of other classification methods. A streamlit application was selected for the interactive presentation of the study results. Among all the models built, XGBoost demonstrated the highest accuracy results, indicating its effectiveness. Therefore, the model can be effectively used as a tool for assessing the correctness of fitness exercises, for example, using a mobile or web app.
Keywords:
FITNESS EXERCISE CLASSIFICATION, MACHINE LEARNING, XGBOOST MODEL, TIME SERIES ANALYSIS, MODELING, CLASSIFICATION QUALITY METRICS
FITNESS EXERCISE CLASSIFICATION, MACHINE LEARNING, XGBOOST MODEL, TIME SERIES ANALYSIS, MODELING, CLASSIFICATION QUALITY METRICS
