Modelling and measuring training information in a network

In this work, we try to answer the question: given a network of observations which are not independent, how can we measure the amount of learning information in it? In particular, given non-independent networked examples, can we find a number n so that we can learn from the networked examples as much as from a dataset of n identically and independently drawn examples? We can answer this question to a large extent by searching an optimum in the space of all possible ways to weight the examples. We can show that we can efficiently perform approximate optimization in this space.status: publishe