If each softmax output neuron represents the probability of the user liking that item, how would you use this model to recommend new posts, as the output neurons do not change? Should I use a totally different system if I am trying to recommend new items? What are the ways I could structure my model so that it can predict new items?