Estimate the skill of players based on past matches.
Accidentally rediscovered the Bradley–Terry model while playing with keras.
Cool thing about the neural net implementation is that it can model the likelihood function non-linearly and, as a bonus, you can even play with latent variables and try to understand what makes winners winners and losers losers. Too bad it needs tons of data though.