[major] lagged regressor with interaction modeling (shared NN)

[karl-richter]

Status quo
For each covariate (aka lagged regressor), a seperate net is trained.

Change
All lagged regressors share a network. In plot_parameters, the attribution of each covariate on the forecast is derived using the Captum model attribution method for deep networks.

[alfonsogarciadecorral]

Hi @karl-richter it looks good to me in terms of model architecture changes

before:

now:

[alfonsogarciadecorral]

Hi @karl-richter @ourownstory

Nice one, I think this is a nice improvement!

A few things from me:

• num_hidden_layers - Maybe we can rename this to num_hidden_layers_lagged_regressors or, if want to share with ar_net, num_hidden_layers_lagged. For example, Im building a shared network for future_regressors and I should be able to use a different configuration.
• Out of curiosity, how come you use that function/formula to create d_hidden if it is not passed by user?
• Would it make sense to pass the neural network architecture as an array? Something like covar_net_layers_array which would replace current n_hidden and num_hidden_layers

[alfonsogarciadecorral]

Hey @karl-richter @ourownstory

The commit I did was about removing num_hidden_layer and d_hidden and instead initialiting the ar_net and covar_net networks through then new variables: ar_net_layers_array and covar_net_layers_array

Have a look, because it implies a few changes:

• We remove all which was related with num_hidden_layer and d_hidden . So now the user either specifies the network or it assumes num_hidden_layers=0. There is no option to pass the num_hidden_layers and then getting the d_hidden through a heuristic method.
• config_model --> this will no longer contain num_hidden_layer and d_hidden and it will have the covar_net_layers_array
• I removed the possibilty of passing the ‘local’ network on method add_lagged_regressor as it is a shared network now.

If you agree on the new changes there are two things missing that I will do as soon as you give thumbs up
IMPORTANT -MISSING:

• Change tutorials documentation to remove the num_hidden_layer and d_hidden
• Add an warn message saying that num_hidden_layer and d_hidden is not used anymore

[ourownstory]

Excellent work @karl-richter and @alfonsogarciadecorral !!

oskar reviewed it once more