The idea of geometric monitoring has been suggested in order to monitor non-linear functions over data collected by distributed sources. The main concept behind the geometric monitoring approach is that each distributed site can safely perform the function monitoring over an appropriate subset of the input domain. In the first part of our talk we first demonstrate how the geometric monitoring approach can be used as a general framework in order to detect ouliers in sensor network applications, for a variety of functions that test the similarity of sensor nodes. Our techniques can manage to detect outliers with 100% accuracy (assuming no message losses), and require the transmission of messages only at a fraction of the epochs, thus allowing nodes to safely refrain from transmitting in many epochs. 

 

In the second part of our talk we examine whether the distributed monitoring mechanism can become more efficient, in terms of the number of communicated messages, by extending the geometric monitoring framework to utilize prediction models. We initially describe a number of local estimators (predictors) that are useful for the applications that we consider and which have already been shown particularly useful in past work. We then demonstrate the feasibility of incorporating predictors in the geometric monitoring framework and show that prediction-based geometric monitoring in fact generalizes the original geometric monitoring framework. We propose a large variety of different prediction-based monitoring models for the distributed threshold monitoring of complex functions.