The v1 implementation of higher-order network models relies on lifting the order from a network up to the desired order $k$, mirroring the multi-order model procedure without caching the lower-order networks. In the context where we have path data, this may be less efficient than simply sliding a length $k$ window over the input data, especially for higher orders and large networks where the line graph transformation may be computationally heavy. The performance tradeoff may depend on whether the user is running on CPU or GPU, since the functions in algorithms.lift_order operate on tensors. This issue is just a reminder that this potential performance tradeoff may be worth investigating in the future.
The v1 implementation of higher-order network models relies on lifting the order from a network up to the desired order$k$ , mirroring the multi-order model procedure without caching the lower-order networks. In the context where we have path data, this may be less efficient than simply sliding a length $k$ window over the input data, especially for higher orders and large networks where the line graph transformation may be computationally heavy. The performance tradeoff may depend on whether the user is running on CPU or GPU, since the functions in
algorithms.lift_orderoperate on tensors. This issue is just a reminder that this potential performance tradeoff may be worth investigating in the future.