When it comes deploying machine learning models on Smartphones processing time is critical.
The approach taken in this study applies LDA-PCA feature extraction to the UCI Human Activity Recognition (HAR) dataset, which contains a large volume of data from a Samsung Galaxy S II smartphone equipped with an accelerometer and gyroscope, and deploys a non-linear, RBF kernel, SVM classifier to decrease processing time and improve classification performance. SVM, LDA, PCA among other modules used throughout this study were implemented in python using scikit-learn.
It was found that mapping the UCI HAR dataset to a higher dimensional space (Case 2 vs Case 1) slightly improves accuracy but increases processing time. We also see in Case 4 that the LDA feature space may be more linearly separable in an infinite vector space relative to Case 3. Lastly, accuracy for the kSVM classifier is highest in Case 5 if we combine features extracted from PCA to the LDA subset. Processing time also decreases 1.05s upon comparing Case 1 to Case 5 which can be of interest for HAR smartphone applications.
