At the beginning of a season, when only a few games have been played, a Bayesian network weighted by starting rankings is used as long as there are whole groups of teams that have not played one another, but once the graph is well-connected, the weights are no longer needed. Sagarin claims that from that point, the rankings are unbiased.