Style Classification Analysis

This repository builds a series of classifiers that predict whether melodies from the Essen corpus originate from China or Europe, using numeric features from the Python package melody-features as predictors. It reproduces the manuscript’s confusion-matrix figures, runs exploratory factor analysis and a factor-based logistic model in R, and benchmarks logistic regression for each feature-extraction source (IDyOM, jSymbolic, etc.).

Run everything from the repo root unless noted otherwise.

---

Prerequisites

Requirement	Notes
Python 3.10+	Check with python3 --version.
R	Check with Rscript --version. Install from CRAN.
melody-features	Installed via requirements.txt. It ships the Essen corpus used to resolve melody paths from basename lists.

---

One-time setup

1. Clone and enter the repo

git clone https://github.com/dmwhyatt/Style-Classification-Analysis.git
cd Style-Classification-Analysis

2. Python environment

python3 -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate
pip install -r requirements.txt

3. R packages (for factor_logistic.R only)

Rscript -e 'install.packages(c("tidyverse", "psych", "jsonlite"), repos="https://cloud.r-project.org")'

ggplot2 is included in tidyverse.

4. melody-features

Feature extraction is run using the Python package defaults. Some inputs may be skipped (e.g. unsupported or polyphonic). This is expected behaviour.

---

Dataset

Two files list melody basenames (no paths):

File	Role
usable_china.txt	One basename per line → pool for China.
usable_europa.txt	One basename per line → pool for Europe.

logistic.py uses every China basename and draws a random subset of Europe (random.seed(42), at most 2200 melodies).

---

Full pipeline

With .venv activated:

python logistic.py
python xgbclassifer.py
Rscript factor_logistic.R
python factor_logistic_plot_confusion.py
python comparison.py

Step	Command	What it does
1	python logistic.py	Builds essen_china_europe_features.csv on first run (this can take a long time due to IDyOM runs). Same stratified train/test and CV as other scripts. Writes Figure 1 (confusion_matrix.pdf) plus coefficient and permutation-importance artefacts.
2	python xgbclassifer.py	Needs the features CSV. Same split/features as 1. Writes Figure 2 (xgb_confusion_matrix_test.pdf).
3	Rscript factor_logistic.R	EFA on the same numeric features (9 factors, promax, parallel analysis). Writes Figure 3 (factor_eigenvalues_elbow.pdf), factor GLM output, and CSVs consumed by 4.
4	python factor_logistic_plot_confusion.py	Reads R’s prediction CSVs. Writes Figure 4 (factor_logistic_confusion_matrix_test.pdf).
5	python comparison.py	Needs the features CSV from 1. Builds or loads source_to_csv_columns_with_novel.json, trains one logistic model per implementation source plus an all features baseline, writes comparison CSV/TeX/PDF and coefficients/*.csv.

First run of Step 1 can take a long time. Later runs load essen_china_europe_features.csv and skip re-extraction unless you delete that file.

---

Figures for main.tex

Figure	Output file	Produced by
1	confusion_matrix.pdf	python logistic.py
2	xgb_confusion_matrix_test.pdf	python xgbclassifer.py
3	factor_eigenvalues_elbow.pdf	Rscript factor_logistic.R
4	factor_logistic_confusion_matrix_test.pdf	python factor_logistic_plot_confusion.py (after 3)

---

Factor network webapp

Rscript factor_logistic.R also writes a self-contained 3D interactive visualization of the eight-factor solution to docs/:

File	Role
docs/index.html	Three.js / 3d-force-graph viewer.
docs/network_data.js	Nodes (factors + variables with |loading| > 0.3) and links.
docs/network_data.json	Same data as a portable JSON sidecar.

Melody examples

python build_melody_examples.py populates docs/melody_examples/ with a piano-roll PNG and a synthesized WAV for the 3 highest and 3 lowest-scoring melodies for every feature node and every factor node in the network. Clicking any node then displays these examples.

• Features are ranked by their value in essen_china_europe_features.csv.
• Factors are ranked by the regression factor scores in factor_scores_for_logreg.csv (produced by factor_logistic.R).
• All of this is precomputed to make the webapp performant.

---

Reproducibility

• Random seeds: 42 is fixed in the Python scripts (train_test_split, CV folds, Europe subsample, XGBoost, etc.) and in factor_logistic.R (set.seed(42)).
• Same train/test rows across logistic.py, comparison.py, and xgbclassifer.py — Keep test_size=0.2 and seeds unchanged.
• Invalidate the feature cache — Delete essen_china_europe_features.csv to force re-extraction (e.g. after changing usable_*.txt or upgrading melody-features in a way that affects columns).