Evolutionary data sampling for user movement classification

Smartphones are nowadays used for recognizing people´s daily activities and habits, by collecting and analysing user activity information in real-time. In order to demonstrate this methodology, we have developed GPSTracker¹ a prototype application for Android phones, which collects position, speed, altitude and time information and performs real-time classification of user´s movement. The GPSTracker application also uses geo-location information abouts Points Of Interest (POIs) such as bus or metro routes, parks and stadiums in order to improve the set of features used for the classification of a type of movement. In this work, we use evolutionary algorithms, in order to reduce the number of samples required for training our classifier, without loosing in classification accuracy. The resulting model, a) is able to provide personalized solutions, tuned to each individual users movement abilities, b) better adapts to unbalanced training data, due to the generation of training samples from the existing ones, c) performs an initial sampling of the training data, which reduces requirements for computational resources and improves the classification performance.