engleski

Inferring Presence Status on Mobile Devices

Automatically inferring presence status on smartphones without user intervention is applicable in many IM and VoIP communication services, but it has not yet been implemented in any of them. It is a problem that requires accurate and energy efficient machine learning methods and a well-chosen subset of features. The objective of this thesis was to test several classification algorithms and to compare them by different criteria: accuracy, training time, test time and the possibility of incremental learning on the given dataset. All the given criteria meet in boosting algorithm AdaBoost based on decision stumps, in fact in its incremental edition called Learn++. C4.5 and 1-NN are also possible candidates, with some necessary adjustments. Several GP variants were tried for a potential use after offline training phase and decision tree based GP showed the highest accuracy, but the slowest training time. The accuracy on accumulated datasets from different users does not decrease significantly. The exhaustive feature selection method was performed and within more than 80% of users only the energy efficient features are necessary to achieve the best possible accuracy.

presence status; mobile computing; machine learning; ensemble learning; genetic programming

nije evidentirano