This repository contains all code written for our research project. Please find our final paper here. Here's a breakdown of all the files that are included:
data.csv is a csv containing all unprocessed data returned by our Qualtrics survey.
likelihood_data.csv and posterior_data.csv contain postprocessed data split between the likelihood (behavior|intent) and posterior (intent|behavior) versions of the survey. Both csvs are organized with each row containing the responses of one survey participant with ratings converted to probabilities. This means that within likelihood_data.csv, all P(action|intention) values for each of Friendship, Relationship, Hookup sum to 1, while within posterior_data.csv, all P(intention| action) values for each potential action (potential text messages you receive) also sum to 1. All data processing and normalization were done in process.ipynb. All rows are anonymized, with demographics of each respondent stored via the ordering from Qualtrics; this ordering is expressed in demographic_dict dictionary in the visualization and calculation notebooks. The columns have been renamed to have more meaningful, human-readable values.
As mentioned, process.ipynb contains all code used to process, normalize, and rekey the raw data before analysis could be performed. visualization.ipynb contains all code and graphs for visualizing the results of our data, including full plots of the likelihood and the posterior across each intention/action, and individual likelihood/posterior plots broken down by demographic. calculation.ipynb contains code written to develop our model to compute priors. We did this by computing the average likelihood and posterior distributions and performing a grid search to check possible values for the prior distribution. We considered models generated by combining the average likelihood distribution and with the potential prior, and we compared the generated posterior to the average observed posterior with KL divergence as our distance function. The prior values with the lowest KL divergence became our optimal prior.