I am trying to learn a model to predict the binary outcome of a computer game. The input data consists of the character picks by each of the ten players (two teams of 5, 150 possibilities each, with removal). This results in one-hot encoded, balanced dataset with 50k rows and +3k cols.
Given the sparsity of the input vector are there any heuristics for the NN design? I've read that Adadrag can be a good choice of optimized for example. On the other hand gridsearching a simple Net has yielded confounding results (all over the place). I was looking to improve on a 54% test accuracy on NBayes and LogReg. Are ff NN just not good a dealing with sparse data?
Any insight appreciated.