The Preprocessing for Predicting of Physical Activity Recognition

Abstract

This experiment examined the preprocessing for predicting of physical activity recognition model to access the relationship between time duration of sensors, the single tri-axial accelerometer, and fitness recognition (sitting, standing, walking, and running). The experimented with sixteen students (62.5% male and 37.5% female, age between eighteen through twenty-three year old) of the Informatics school at Walailak University. The authors had the experimental setup with the split dataset, 80% for training and testing, and 20% for validation, and repeated k-fold Cross-Validation (number=10, repeats=3) for resampling method to evaluate model performance for baseline models. When the authors measured model’s performance, the authors found the follows results. First – the raw dataset with 123,156 samples, the best models performance has accuracy level with KNN: k-Nearest Neighbor and RF: Random Forest is 100%. Second – the aggregate dataset time duration 1 second with 1,240 samples, the best models performance has accuracy level with RF: Random Forest is 100%.Third – the aggregate dataset time duration 5 seconds with 251 samples, the best models performance has accuracy level with RF: Random Forest is 99.5%. Fourth – the aggregate dataset time duration 10 seconds with 128 samples, the best models performance has accuracy level with KNN: k-Nearest Neighbor is 96.82%. Fifth – the aggregate dataset time duration 15 seconds with 86 samples, the best models performance has accuracy level with KNN: k-Nearest Neighbor is 96. 21%.Sixth – the aggregate dataset time duration 20 seconds with 66 samples, the best models performance has accuracy level with LDA: Linear Discriminant Analysis is 98%.Seventh – the aggregate dataset time duration 25 seconds with 54 samples, the best models performance has accuracy level with KNN: k-Nearest Neighbor is 96.33%. Moreover, finally, Eight – the aggregate dataset time duration 30 seconds with 46 samples, the best models performance has accuracy level with KNN: k-Nearest Neighbor is 93.61%. In the future work, the authors planned to get more accuracy model by adding more features from another sensor, heart rate. Mining data collected from sensors provide valuable result in the physical activity recognition area. The improvement in performance is required especially in the healthcare field. The more increasing of using the wearable device, the wider opportunity in the data mining research area can be.