Lightweight adaptation to situational changes in classifiers of multimodal human data multimodal human data: Dissertation

Intelligent computer applications need to adapt their behaviour to contexts and users, but conventional methods to train multimodal classifiers do not suit to this purpose because they require acquiring large sets of labelled data for each situation. Due to large variety of usage contexts of personal applications, no developer can predict all these situations, to say nothing of collecting adequate training databases for them. Hence personal applications require new methods for adapting to changing runtime contexts. As runtime adaptation largely relies on interaction with end users, these methods should be fairly lightweight with respect to standard ones, i.e. they should require much less domain knowledge and explicitly acquired data.This thesis introduces lightweight solutions for adapting reasoning models to situations at runtime, identifies important context and application characteristics and provides guidelines for considering these factors in adaptation design. The proposed solutions have been validated experimentally with realistic data sets, and the results have confirmed that they considerably reduce the dependence of context- and user-adaptive classifiers on domain knowledge and explicit interaction efforts. Studies with personal assistive applications have also demonstrated that users can accept the proposed lightweight adaptation even when its accuracy is relatively low.