The thesis describes the challenge of determining optimal moments for delivering notifications to mobile devices. Traditional machine learning models, which use data on user location, activity, and device status, can effectively predict these moments but require data to be sent to the cloud, which compromises privacy. In this thesis, we use federated learning, which enables model building in a distributed way without sending personal data to a server. We developed an Android application using the TensorFlow Lite and Flower frameworks, enabling on-device learning. Testing on real mobile devices demonstrated that federated learning allows for effective and privacy-preserving prediction of user interruptibility.
|
