One of the fundamental assumptions for machine learning based text classification systems is that the underlying distribution from which the set of labeled-text is drawn is identical to the distribution from which the text-to-be-labeled is drawn. However, in live news aggregation sites, this assumption is rarely correct. Instead, the events and topics discussed in news stories dramatically change over time. Rather than ignoring this phenomenon, we attempt to explicitly model the transitions of news stories and classifications over time to label stories that may be acquired months after the initial examples are labeled. We test our system, based on efficiently propagating labels in time-based graphs, with recently published news stories collected over an eighty day period. Experiments presented in this paper include the use of training labels from each story within the first several days of gathering stories, to using a single story as a label.