To secure location privacy of social networks users in a manageable way, a large body of existing works focuses on automating the location-sharing decisions with online social contacts by defining machine-learning based location sharing policies. These policies are learned over actual user-responses to information sharing requests from their social connections under varying contextual conditions (e.g., time of day, location, etc.) and replicate user behavior autonomously for subsequent requests.

However, training high-performance location sharing policies, e.g., those with high accuracies, typically impose high computational demands. Hence training high-performance policies on mobile devices, which are computationally resource-constrained, is very challenging. In this regard, we propose a novel policy learning methodology whereby high-performance policies can be trained with low computational investment. This computational gain of our methodology is based on the selection of high quality training data. We also show that the high quality training data is composed of those training examples that are associated with strong mobility routine of the users. Thus we also propose an algorithm for efficient estimation of the dynamic strength of movement routine on mobile devices. Our evaluations on real world data demonstrate the usefulness of our approach.