liketrainer

Code for the 'Like trainer, like bot? Inheritance of bias in algorithmic content moderation' study, presented at SocInfo 2017

Paper available on Arxiv

Training data

Data used in the study is taken from previous work by Wulczyn et al, and can be found here

Building classifiers

The basic classifier (using all the training data) is built with make_clf.py. Male-only, female-only and mixed-gender classifiers are labelled accordingly.

make_models.py builds 10 classifiers. In order to generate random samples that are reproducible, the numpy random seed function is used. The resulting classifiers are named 1-10 after the random seed used to generate the sample on which they were trained.

coefficients.py extracts the coefficients from a set of classifiers.

Building test data

The test dataset used is test_detox.csv and is generated with make_mixed_test.py.

Results

The results of the main tests are in test_results_balanced.csv.