austincap / alwayssunny

Various tools for analyzing Always Sunny

AlwaysSunny

Some files and scripts needed for analyzing Always Sunny

The purpose of this pet project is to see if there are any patterns in the data I personally extracted from each episode of Always Sunny in Philadelphia, specifically: what makes a good episode? Data is tragically underutilized in the art world. While watching the show I noticed that the writers seemed to emphasize the fact that the core group splits up in different ways, and that in some episodes there seem to be a clear "winner." Ultimately this sort of "character grouping" analysis is applicable to all episodic, character-driven TV-shows.

I started with this data coded by hand for fun and converted everything (again by hand but not as fun) to both episode production codes and episode codes used by imdb.

Then I turned that data into a dictionary of matrices to represent which character grouped with who and whether or not that team "won" for each episode using this script. I tried running it on a simple neural net with IMDB ratings for the label data just to see what happened (nothing useful).

I wanted to make some Tableau viz's of this but I couldn't think of a good way to convert my data into a MySQL database. Fortunately I found Simba's Mongo->SQL->Tableau conversion drivers. I consolidated my raw data and imdb ratings in to a mongo database with this script.

So far the most interesting thing all this work has produced is this image I made in Tableau that shows what groups made for the best episodes.

alwayssunny's Issues

Unsupervised learning

Hey there!

I actually had the same idea for doing analysis on Always Sunny episodes but stalled out almost immediately in the data collection process. One of the things that I thought would've been really cool would be to track the "allegiances" of the gang to see who was pairing up with whom and who was the most likely to double cross each other. That proved to be really hard to keep track of because what they were aligning themselves over was changing over the course of the episode.

Have you thought of trying to do any unsupervised learning on your data such as kmeans or NMF? Could be cool to see how the episodes group together based on what is happening. Anyway, I'm glad someone had the initiative to get past the data collection part and make some visualizations.

I also saw you manually coded up a neural network, have you looked into something like Keras? It would run a lot faster, but the amount of data is still woefully small to be using a net with.

Best,
Erich