One of the primary challenges is that such data includes both numeric features (e.g. box stats) and contextual labels (e.g. teams), making each observation both relative and coupled.
Perhaps the most natural representation of such data is a time-dependent graph, where the comparisons represent edges between two nodes (competitors). Unfortunately, this kind of graph structure appears to preclude naive applications of ANNs developed for tabular and sequence data (e.g. LSTMs, transformers, etc).
Does anyone know of any good references that demonstrate how to account for the graph structure of such problems? Is it a terrible idea to essentially one-hot encode the adjacency matrix and provide it as an input to an LSTM or Transformer architecture? Are there more efficient ways to encode paired comparisons?
I tried something like a neural network version of Elo with more features (i.e. RNN), but I suspect I was suffering from exploding/vanishing gradients since I was chaining the same hidden layer update step for thousands of games.
Any advice or direction here would be much appreciated!