Feature selection on Human Activity Recognition dataset using Minimum Redundancy Maximum Relevance

Human Activity Recognition is a research field that aims to identify the activities carried out by a person. Recognition can be done by using information that is retrieved from various sources, for example: using inertial sensors. Lately, many people use smartphone with built-in inertial sensors such as accelerometer and gyroscope which make these smartphone capable of recognizing human activities. But, an optimization must be done to minimize the computational process of the recognition system so that the system could be function properly in smartphone that has limited processing power. In this study optimization was done by reducing the numbers of features used in the dataset. Using an available public human activity recognition dataset, mRMR (Minimum Redundancy Maximum Relevance) feature selection method was applied to the dataset to reduce the numbers of features. Results from this study show that using mRMR method in dataset 7∶3, the numbers of features can be reduced from 561 features to 201 features, while still maintaining the accuracy at 95.15% using SVM classifier and 94.23% using MLP classifier. While using mRMR method in dataset 8∶2, the numbers of features can be reduced from 561 features to 154 features with the accuracy 95.18% using SVM and the numbers of features can be reduced from 561 to 211 with the accuracy 94.84% using MLP. Both dataset using the same threshold that is 95% for SVM and 94% for MLP. For the running time computation, in dataset 7∶3 the running time becomes 42.57% from the initial running time using SVM and 12.63% from the initial running time using MLP. While in dataset 8∶2 the running time becomes 60.5% from the initial running time using SVM and 14% from the initial running time using MLP.