Develop potential outcome graphical model with preprocessing as intervention

[scheurich-sarah]

Preprocessing as a binary treatment: you have two treatments and we can observe the counterfactual by applying both kinds of preprocessing to the data

treatment 0 (control): processed with word2vec

• Y_0 or Y(0) potential outcome if processed with word2vec
  treatment 1: processed with doc2vec
• Y_1 or Y(1) potential outcome if processed with doc2vec

Causal effect evaluates what would happen if you use doc2vec instead of word2vec
For any individual requirement, we can observe

• Does it accurately trace the requirement regardless of which preproc you use, so No effect
• Does it inaccurately trace the requirement regardless of which preproc you use, so No effect
• Traces requirement with word2vec but not doc2vec, so Negative effect
• Traces requirement with doc2vec but not word2vec, soPositive effect

We compute the average of this result for every requirement
ACE=E[Y_1 ]−E[Y_0 ]

#109