modeling_citations

Motivation

This project is designed to recreate the bias in citation practices in neuroscience reported here. Here, the perceived gender of authors is inferred by comparing the first names of authors to a database of names from multiple countries. It is important to note the aim is the understand the bias associated with the perceived gender of a name, not the individuals true gender identity or sex. This work shows (1) the men-led teams are cited more than women-led teams given their presence in the field (2) that these disparities are increasing over time, despite the field becomng more diverse, and (3) that these disparities are driven mostly by the majority (men-led teams). Here, we seek to gain an understanding of the possible drivers of and most effective interventions for this phenomenon through simulations of citation practices.

We define artificial agents (authors) that have different mental estimates of the network of people that make up the field of neuroscience. In this network, nodes are authors and edges are co-authorships. Citation lists are generated from biased random walks on these networks. These estimated networks can be influenced by meetings with other authors. Estimated networks will also periodically prune (or forget) authors who do not frequenctly appear in citation lists.

Dependencies

This project is done entirely in Python (v>=3.6), and the package versions used are given in the requirements.txt file

Code

Classes

1. Author: defined in functions/author_fns.py. This class defines the properties and functions of 'authors' in the simulation. Each author has a gender, and 5 parameters that determine how it interacts with other authors.

Key Functions

1. Author.init_network(): defined in functions/author_fns.py.This function intiaites an Authors estimate of the field, using a Levy flight on the true co-author network/ Can be biased towards authors of one gender with net_bias
2. Author.get_cites(): defined in functions/author_fns.py. Generates a citation list from a random walk an the authors network. Can be biased towards one of the genders with walk_bias
3. Author.forget(): defined in functions/author_fns.py. Removed some number of authors that have not been viewed recently. The number is determined by forget_bias
4. compare_nets(): defined in functions/author_fns.py. Decides whether two authors will mett with and learn from each other based on their meet_bias parameters

Scripts

1. eda_preproc.ipynb: this script shows basic statistics (degree distribution, number of nodes, etc.) for the neuroscience co-author network used in this research
2. class_fn_test.ipynb: this script tests failure cases for all functions, and plots how basic statics (network size, gender distribution) change for different values of bias parameters
3. simulations.ipynb: this script simulates meetings between authors, and plots citation practices over simulations