by Svitlana Glibova

photo by cottonbro from Pexels

Stepmania is an open-source software designed to support many styles of rhythm game play, but one of its most widely-used formats is for 4-panel “dance games” (like Dance Dance Revolution or In The Groove, but not licensed). The software was released in 2001 and has since been updated through version 5.3 - the concept behind this software is to allow players to play a “DDR-like” game with virtually boundless customizations, including custom themes, statistics, and charts (arrow patterns for songs).

There is not currently a standardized scale or model for determining chart difficulty being used within the community, leaving the determination up to step-artist and community input. While extensive domain experience means that chart ratings are relatively consistent, there is variance that can be addressed.

The problem that I aim to solve is using exploratory data analysis and predictive modeling to help establish a standardization for song chart difficulty ratings based on various features of the chart, such as BPM, number of steps, and counts of technical elements.

Creating a method of standardization will help examine which features are the most important in determining difficulty and help step-artists remove the intuitive/guesswork element from rating their own charts.

itgpacks.com is a community-maintained spreadsheet containing the majority of song packs released in the last five years. The majority of this data came from here, with the exception of a few packs downloaded from other sources.

The data is extracted from each individual song file using a parser written in Python to collect the features used for EDA and predictive modeling. A huge, huge, huge thank you to Tim Murphy, a friend and dance game player, for writing the code for this parser. You can check out their GitHub here..

The data is stored in data in two folders: stam.csv and not_stam.csv. Please keep the folder structure the same way for this notebook to work. If you would like to include your own song files, you can visit this Github Repo for the code to run a parser over your own song files. To use:

python src/python/step_parser.py path/to/your/songfiles path/to/data/yourfile.csv

Make sure you have the .csv file in data and in the notebook, change the path to look like this:

your_dataframe = pd.read_csv('yourfile.csv')

The rest of the notebook must then use the your_dataframe variable from here on out.

Here you can find a full list of each generated feature and a a brief description.

Distribution of tech data difficulties:

The difficulty values here have a smaller distribution than that of the stamina data below. For this reason, it seemed important to preserve the distinction with two datasets.
Relationship of step count vs rating for tech data:

There is a much more linear relationship with the tech data than with stamina because stamina has a much broader step count distribution as well.

Stamina difficulty distribution:

Relationship of step count vs. rating for stamina data (note that there are very few songs in the upper difficulties, which skewed the correlation):

Using simple linear regression on scaled data as a first simple model was surprisingly successful, as the residual plot above illustrates.
R-squared values hovered around .95, but a few of the residuals (true value minus predicted value) were too large to be considered accurate.

Residuals for tech data:

Residuals for stamina data:

Using ridge regression with a lower regularization values (alpha = .1 to alpha = .3) increased the model's R-squared as well as facilitated feature extraction.
With scaling, power transform, and ridge regression, the mean R-squared was approximately .962 and below are the largest coefficients extracted from the data.
Tech:
step_count
song_nps
nps_per_measure_max
nps_per_measure_avg - surprisingly, this coefficient was negatively correlated.
nps_per_measure_std

Stamina:
step_count
measure_count - negative correlation, but potentially explained as a relationship to density as challenging songs tend to be more dense but aren't necessarily longer.
song_nps
nps_per_measure_max
nps_per_measure_std
stream_total

Using the XGBoost regression algorithm alone drastically increased the training R-squared scores to over .99, but showed some overfitting with the validation set.

Feature extraction proved to be slightly more challenging with this model but NPS still proved to be consistently important across both datasets.

Technical:

Train R^2: 0.9905622716862734
Test R^2: 0.9682993868229007
Train RMSE: 0.3392237128757486
Test RMSE: 0.6358308816629418

Stamina:

Train R^2: 0.9921925792873305
Test R^2: 0.9792039976085652
Train RMSE: 0.3401577104917719
Test RMSE: 0.5312995167797424

The final model was able to accurately predict on validation data with minimal error, although it looks like there was some challenge with higher-level difficulties for stamina data (most likely due to there being fewer data in this realm), and due to the variance of the technical data.
There was some slight overfitting with both datasets, but the R-Squared values did not differ significantly from training to test sets.

To continue this project, I aim to collect more song data and potentially re-train the model with more information, although the key features are unlikely to change. I also intend to tune a classifier model beyond the initial EDA stage.

The web application of this model can be found in the below section, and the next step will be to allow the user to upload their own .sm or .ssc files, rather than having to parse the data separately before uploading the .csv.

The first iteration of a web app to predict user-input data is live here! The code can be found here.

├── data
  ├── feature_dictionary.txt
  ├── not_stam.csv
  └── stam.csv
├── img     
├── notebooks  
  ├──eda_notebook.ipynb
  └──model_testing.ipynb 
├── src
  ├── models
    ├── tech_model.pkl
    └── stam_model.pkl
  ├── visualizer.py   
  └── data_cleaning.py  
├── final_notebook.ipynb  
├── README.md
├── sm_model_env.yml
└── presentation.pdf - contains a non-technical presentation of the findings 

A huge thank you to the creators of Stepmania and to quietly-turning and hurtpiggypig for the Simply Love Theme for creating and maintaining this software, as well as all contributing stepartists for the data used in this project.