My goal is to build a recommendation model for which I want to use Neural Network (LSTM). The user will give some input keywords and the model should return the suggestions (classes) based on relevance to the input keywords. My dataset contains few thousand documents and each document has exactly one class, so 1000 documents = 1000 classses. Every document contains 50-800 words.
My question is if it is a good idea to train a neural network on such data where there is only one instance of each class? If not then how can I achieve my goal.
I have already tried Naive Bayes and it works efficiently with such data. Any help would be appreciated.