Performance Evaluation of Shadow Features as a Data Preprocessing Method in Data Mining for Human Activities Recognitions

Abstract

A Human Activity Recognition (HAR) classification model is used to predict the class or predefined type of human activity. With the limited amount of available original features of human activity, the classification performance is usually mediocre. One solution is to enrich the information of the original data attributes. The objective of this study is to find a suitable feature transformation method for inducing an accurate classifier for HAR. A novel concept for enriching the feature information of HAR is called Shadow Feature. Two versions of Shadow Features are implemented here. They are being tested via RapidMiner to see which version is more suitable for HAR. The experiment results show that the four data pre-processing strategies could be ranked by their performance as follow: shadow feature 2 > shadow feature 1 > statistical features > original features. Algorithm-wise, ensemble algorithms are able to improve the HAR classification performance while a single decision tree is shown to be a weak classifier. Finally, it is observed that good performance can be achieved when shadow features are applied over datasets of drastic activity; in this case shadow feature 2 is better than shadow feature 1. For datasets of subtle activity shadow features do have advantages too, though slightly; in this case shadow feature 1 works better than shadow feature 2.