Nondestructive active seismic sources have been successfully used to image and monitor the earth on many different scales. Although the raw signals generated by these sources can be weak, the signals can be retrieved over great distances using seismic interferometry and stacking. When using these methods, a major limiting factor on range and signal‐to‐noise ratio of the retrieved signal is the presence of persistent noise sources in the same frequency range as the signal from the active source. In this article, we will introduce a method to separate a desired active source interferometric signal from coherent background noise sources via supervised machine learning, based on a sparsely weighted semi‐nonnegative matrix factorization approach. We apply this method to data recorded in an active underground mine, where a nondestructive pneumatically powered active source is used to try to track the growth of a sublevel cave. We show that the method is able to extract the signal from the active source on a sensor more than 1 km away, while normal seismic interferometry and stacking fails.