When an underwater acoustic sensor network (UASN) is applied to underwater data collection, different data importance rating (DIR) of sensor nodes will affect the scheduling time slot of data collection. In this paper, we propose a Q-learning and DIRbased media access control (Q-DIR MAC) protocol for dynamic clustering underwater acoustic sensor networks (UASNs), in which the nodes in the network may drift with the movement of ocean currents. We use k-mean algorithm to divide the nodes into several clusters. Each partitioned cluster is composed of one cluster head (CH) and several cluster members (CMs). The CMs can be divided into three levels according to the DIR: non-urgent, normal, and very urgent. The number of three types of nodes follows normal distribution. The data importance of each node is introduced into reward function design of Q-learning. The results show that, in the dynamic clustering UASNs, the proposed QDIR MAC protocol can ensure that important data can be sent to the destination node in time without reducing the data success rate under the condition of priority transmission mechanism.