Packet losses tend to occur during short error bursts separated by long periods of relatively error-free transmission. There is also a significant spatial correlation in loss among the receiver nodes in a multicast session. To recover packet transmission errors at the transport layer, tree-based protocols construct a logical tree for error recovery before data transmission is started. The current tree construction scheme does not scale well because it overloads the sender node. We propose a scalable recovery tree construction scheme considering these properties. Unlike the existing tree construction schemes, our scheme distributes some tasks normally handled by the sender node to specific nodes acting as repair node distributors. It also allows receiver nodes to adaptively re-select their repair node when they experience unacceptable error recovery delay. Simulation results show that our scheme constructs the logical tree with reduced message and time overhead. Our analysis also indicates that it provides fast error recovery, since it can reduce the number of additional retransmissions from its upstream repair nodes or sender node.