Multiple adaptive mechanisms for predictive models on streaming data.
Conference: Bournemouth University, Faculty of Science and TechnologyAbstract:
Making predictions on non-stationary streaming data remains a challenge in many application areas. Changes in data may cause a decrease in predictive accuracy, which in a streaming setting require a prompt response. In recent years many adaptive predictive models have been proposed for dealing with these issues. Most of these methods use more than one adaptive mechanism, deploying all of them at the same time at regular intervals or in some other fixed manner. However, this manner is often determined in an ad-hoc way, as the effects of adaptive mechanisms are largely unexplored. This thesis therefore investigates different aspects of adaptation with multiple adaptive mechanisms with the aim to increase knowledge in the area, and propose heuristic approaches for more accurate adaptive predictive models. This is done by systematising and formalising the “adaptive mechanism” notion, proposing a categorisation of adaptive mechanisms and a metric to measure their usefulness, comparing the results after deployment of different orders of adaptive mechanisms during the run of the predictive method, and suggesting techniques on how to select the most appropriate adaptive mechanisms.
The literature review suggests that during the prediction process, adaptive mechanisms are selected to be deployed in a certain order which is usually fixed beforehand at the design time of the algorithm. For this reason, it was investigated whether changing the selection method for the adaptive mechanisms significantly affects predictive accuracy and whether there are certain deployment orders which provide better results than others. Commonly used adaptive mechanism selection methods are then examined and new methods are proposed.
A novel regression ensemble method which uses several common adaptive mechanisms has been developed to be used as a vehicle for the experimentation. The predictive accuracy and behaviour of adaptive mechanisms while predicting on different real world datasets from the process industry were analysed. Empirical results suggest that different selection of adaptive mechanisms result in significantly different performance. It has been found that while some adaptive mechanisms adapt the predictive model better than others, there is none which is the best at all times. Finally, flexible orders of adaptive mechanisms generated using the proposed selection techniques often result in significantly more accurate models than fixed orders commonly used in literature.