Moebius_converter adds many value 0 interactions during conversion

[landonbutler]

During the conversion from a sparse Möbius representation to another index, sometimes many value 0 interactions are added to the returned InteractionValues.

This is because interactions are added to the interaction_dict, even if their coefficient is 0.

• shapiq/shapiq/game_theory/moebius_converter.py

  Lines 125 to 130 in f3e9e7f

  	val_distributed = distribution_weights[moebius_size, len(interaction)]
  	# Check if Möbius value is distributed onto this interaction
  	if interaction in base_interaction_dict:
  	base_interaction_dict[interaction] += moebius_val * val_distributed
  	else:
  	base_interaction_dict[interaction] = moebius_val * val_distributed

• shapiq/shapiq/game_theory/moebius_converter.py

  Lines 199 to 204 in f3e9e7f

  	val_distributed = distribution_weights[moebius_size, len(interaction)]
  	# Check if Möbius value is distributed onto this interaction
  	if interaction in stii_dict:
  	stii_dict[interaction] += moebius_val * val_distributed
  	else:
  	stii_dict[interaction] = moebius_val * val_distributed

• shapiq/shapiq/game_theory/moebius_converter.py

  Lines 275 to 280 in f3e9e7f

  	val_distributed = distribution_weights[moebius_size, len(interaction)]
  	# Check if Möbius value is distributed onto this interaction
  	if interaction in fii_dict:
  	fii_dict[interaction] += moebius_val * val_distributed
  	else:
  	fii_dict[interaction] = moebius_val * val_distributed

Alternatively, the returned InteractionValue could be sparsified upon return.

[mmschlk]

@landonbutler, thanks for pointing this out! If moebius_val * val_distributed is zero do not need to be added to the interaction dictionaries. You are right! If this is not not too much of a problem currently, you could take a look at InteractionValues.sparsify() which does this (however only post hoc and requiring another loop over the interactions.

Nevertheless, the creation could still be optimized further and I will keep track of this here. :)