Enhancing Individual Location Prediction Using Group Data
PhD student at Southampton University
DEI - 3A Room
November 11th, 2011
Better smart home technologies require improved daily location prediction. We propose a new method for prediction of daily life location that is more accurate and requires shorter training periods than the existing state of the art. In more detail, current methods typically assume no prior knowledge about the habits of people before training begins. Where prior knowledge is assumed, it is hand coded, which is inflexible and can be costly. However, we believe that by grouping users based on their habitual behaviours 1) the accuracy of individual location prediction goes up, and, 2) the training time required for new individuals goes down. Specifically, we offer a new approach that harnesses behavioural groupings to refine daily predictions for individual users. This approach is based on the view of behaviour as a mixture model, in which each period of behaviour is selected with fixed probabilities from a set of distributions. In benchmarks against existing approaches using real location data from the MIT Reality Mining dataset we show that this new method improves by 14.3% the average accuracy in the first 10 days of use in daily life prediction, and performs 7.0% better when tested with the most atypical days for the subject (i.e. under data sparsity conditions).
James McInerney received a degree in Computer Science from Oxford University, and a Masters in Artificial Intelligence from Imperial College London. Since January 2011, he has been a PhD student at Southampton University under the supervision of Nick Jennings and Alex Rogers. His research interests are in machine learning and behaviour prediction and analysis. He is working as part of Orchid, a multi-disciplinary project investigating the interaction between humans and computer agents.
Artificial intelligence, robotics and computer vision