In this paper, we consider the problem of user scheduling and pilot assignment in TDD multicell multiuser Massive MIMO systems. While in TDD systems the channel is acquired using uplink pilots, we propose a scheme that utilizes additional downlink probing in order to improve the spectral efficiency. The idea is to dynamically assign mobile users to different clusters based on the directions of their channels through the use of downlink reference beams. This will result in forcing the interference to be centered in semi-orthogonal subspaces without the need for important feedback and therefore enabling reduction of the pilot contamination effect. The scheduled users in each cluster employs orthogonal training sequences with a pilot reuse factor of 1 among the clusters. We then propose a graphical framework for pilot assignment. We show that this problem can be modeled as a max cut problem and we provide an approximation algorithm that optimizes the pilot allocation.