One reason for this is that it improves label convergence time. If you have two PE routers that are separated by three or four hops for example, and there is network instability, the targeted LDP packets will still be able to be exchanged between the two PE routers via alternate routes within the network (assuming dynamic routing and redundancy are part of the network design), even if some links may have been compromised. In this way, the adjacency session will be maintained because the targeted LDP packets found outer routes to reach the two PEs. If targeted LDP sessions were not used, and network instability ensued, adjacencies would be lost and would have to be re-established, causing some network downtime.
BGP can be configured to function in a similar way, when using multi-hop, where neighbour adjacencies can be created between two non-directly connected routers.
I hope this has been helpful!