Subreddit Analysis

This repo includes tools for Subreddit analysis, originally developed for our class project of PSYC 626 in USC, titled “Powered by the Will?: Themes in online discussions of Fitness”.

Installation and Requirements

You need to use Python 3.9, R 4.1.0 and git basically to run the scripts provided in this repo. For Ubuntu, to install essential dependencies:

sudo apt update
sudo apt install git python3.9 python3-pip
pip3 install virtualenv

Now clone this repo:

git clone https://github.com/gchochla/subreddit-analysis
cd subreddit-analysis

Create and activate a python environment to download the python requirements for the scripts:

~/.local/bin/virtualenv .venv
source .venv/bin/activate
pip install .

Usage

Download a subreddit into a JSON that preserves the hierarchical structure of the posts by running:

python subreddit_analysis/subreddit_forest.py -r <SUBREDDIT_NAME>

where <SUBREDDIT_NAME> is the name of the subreddit after r/. You can also limit the number of submissions returned by setting -l <LIMIT>. The result can be found in the file <SUBREDDIT_NAME>-<NUM_OF_POSTS>-pushshift.json.

Transform this JSON to a rectangle (CSV), you can use:

python subreddit_analysis/json_forest_to_csv.py -fn <SUBREDDIT_NAME>-<NUM_OF_POSTS>-pushshift.json

which creates <SUBREDDIT_NAME>-<NUM_OF_POSTS>-pushshift.csv.

To have a background corpus for control, you can download posts from the redditors that have posted in your desired subreddit from other subreddits:

python subreddit_analysis/user_baseline.py -fn <SUBREDDIT_NAME>-<NUM_OF_POSTS>-pushshift.json -pl 200

where -pl specifies the number of posts per redditor to fetch (before filtering the desired subreddit). The file is saved as <SUBREDDIT_NAME>-<NUM_OF_POSTS>-pushshift-baseline-<pl>.json

Transform that as well to a CSV:

python subreddit_analysis/json_baseline_to_csv.py -fn <SUBREDDIT_NAME>-<NUM_OF_POSTS>-pushshift-baseline-<pl>.json

which creates <SUBREDDIT_NAME>-<NUM_OF_POSTS>-pushshift-baseline-<pl>.csv.

Create a folder, <ROOT>, move the subreddit CSV to it, and create another folder inside it named dictionaries that includes a file (note: the filename — with a possible extensions — will be used as the header of the loading) per distributed dictionary with space-separated words:

positive joy happy excited

Tokenize CSVs using the r_scripts.
Compute each post’s loadings and write it into the CSV:

python subreddit_analysis/submission_loadings.py -d <ROOT> -doc <CSV_FILENAME>

where <CSV_FILENAME> is relative to <ROOT>.

If annotations are available, which should be in a CSV with (at least) a column for the labels themselves and the ID of the post with a post_id header, you can use these to design a data-driven distributed dictionary. You can first train an RNN to create another annotation file with a predicted label for each post with:

python subreddit_analysis/rnn.py --doc_filename <SUBREDDIT_CSV> --label_filename <ANNOTATION_CSV> --label_column <LABEL_HEADER_1> <LABEL_HEADER_2> ... <LABEL_HEADER_N> --out_filename <NEW_ANNOTATION_CSV>

where you can provide multiple labels for multitasking, thought the model provides predictions only for the first specified label for now. Finally, if annotations are ordinal, you can get learned coefficients from Ridge Regression for each word in the vocabulary of all posts (in descending order of importance) using a tf-idf model to represent each document using:

python subreddit_analysis/bow_model.py --doc_filename <SUBREDDIT_CSV> --label_filename <ANY_ANNOTATION_CSV> --label_column <LABEL_HEADER> --out_filename <IMPORTANCE_CSV>