How to examine effect of variable not used in training a neural network

Question

I am currently using tensorflow to create a neural network that does basic binary classification, and I would like to examine the bias of the model after training.

I have a dataset of about 300,000 samples, and each sample has a list of 112 features. Prior to training, we've selected 20 features that we want to use for prediction - race is a feature in the dataset, but it is not used for training the neural network.

What I'd like is to be able to see how often a variable with a different values for "race" gets misclassified, without including the race feature in the training dataset. Is there a way to essentially drop that feature during training, then re-append it to the dataset to see what the accuracy is on specific groups? Or to tell tensorflow to ignore that variable while training?

Answer 1

If you have

• df the dataset with all the columns
• df_train and df_test the datasets for train and test with all the columns.
• X_train and X_test the datasets for train and test with only the selected columns.
• y_train and y_test the target variables.
• model your model.

you can do something like

# make predictions
model.fit(X_train, y_train)
predictions = model.predict(X_test)

# create a dataframe with predictions and race
# here you can add as many columns as desired
race = df_test['race'].values
results = pd.DataFrame(
        {'predictions': predictions,
         'race': race
        }
)

# inspect the effect of race in predictions - even if race is not used in the model
results.groupby('race').describe()

Also, if you want to go deeper in the field of AI fairness I recommend you this project: https://fairlearn.org/