Using random walk sampling methods for feature learning on networks, we develop a method for generating low-dimensional node embeddings for directed graphs and identifying transition states of stochastic chemical reacting systems. We modified objective functions adopted in existing random walk based network embedding methods to handle directed graphs and neighbors of different degrees. Through optimization via gradient ascent, we embed the weighted graph vertices into a low-dimensional vector space Rd while preserving the neighborhood of each node. We then demonstrate the effectiveness of the method on dimension reduction through several examples regarding identification of transition states of chemical reactions, especially for entropic systems.