Analyzing socially aware movement representations is growing tremendously in the recent developments of multi-agent situations like forecasting human movement and exploration of robot antics on roads along with people in a socially safe environment. In spite of good developments, the present techniques obtained using deep learning, i.e., neural networks find it difficult to generalize the trajectory collisions. This problem mainly results from the non-identical and non-independent type of problem of path forecasting coupled with badly distributed data used for training. Accordingly, if the source of data used for training just keeps coming from the behavior of human beings in situations where the chances of edge cases like collisions are negligible, then making models know the concept of edge case scenarios such as collisions becomes really ambiguous. Through this research work, we target to solve the problem by explicit training edge case or negative scenarios by own control: (i) Defining a social contrasting loss, this regulates the movement by distinguishing real true scenarios having less collision probability from deliberately designed negative samples. (ii) From our previous experience edge case scenarios are specifically designed considering dangerous situations. By doing this, we observe that it significantly decreases the rate of collision of our path/trajectory forecasting.