Context changes detection by one-class svms
For a system that aims at taking into account the user, we need to consider that there are many different behaviors as well as many different users. Hence we need adaptative, unsupervised (or semi-supervised) learning methods. Our idea is to take advantage of wearable computers and wearable sensors (indeed their use is realistic at least for certain categories of people, such as pilots) to retrieve the current context of the user. Wearable sensors can be physiological (EMG, ECG, blood volume pressure...) or physical (accelerometers, microphone...). Contexts are depending on the application using the system and can be behaviors, affective states, combinations of these. Since this problem of context retrieval is very complex, we choose to detect changes at first place instead of labeling directly. Indeed this way we can apply unsupervised and fast methods which saves time for labeling (the labeling task is then applied only when changes are detected). Our interest lies in low level treatments and we present a non parametric change detection algorithm. This algorithm is meant to provide sequences of unlabeled contexts to be analyzed to higher level applications. Detection is made from signals given by non invasive sensors the user is wearing. Note that the methods presented here could as well be adapted to external sensors.
Author: Gaëlle Loosli, National Institute of Applied Sciences