When a human is interacting physically with a robot to accomplish a task, his/her posture is inevitably influenced by the robot movement. Since the human is not controllable, an active robot imposing a collaborative trajectory should predict the most likely human posture. This prediction should consider individual differences and preferences of movement execution, and it is necessary to evaluate the impact of the robot's action from the point of view of ergonomics. Here, we propose a method to predict, in probabilistic terms, the human postures of an individual for a given robot trajectory executed in a collaborative scenario. We formalize the problem as the prediction of the human joints velocity given the current posture and robot end-effector velocity. Previous approaches to solve this problem relied on the inverse kinematics, but did not consider the human body redundancy nor the kinematic constraints imposed by the physical collaboration, nor any prior observations of the human movement execution. We propose a data-driven approach that addresses these limits. The key idea of our algorithm is to learn the distribution of the null space of the Jacobian and the weights of the weighted pseudo-inverse from demonstrated human movements: both carry information about human postural preferences, to leverage redundancy and ensure that the predicted posture will be coherent with the endeffector position. We show in a simulated toy problem and on real human-robot interaction data that our method outperforms model-based inverse kinematics prediction, sample-based prediction and regression methods that do not consider geometric constraints. Our method is validated on a a collaboration scenario with a human interacting physically with the Franka robot.