# Music Opinion Dynamics
<img width="1774" height="887" alt="d1381d7a-6e4c-48fb-83c2-dacd0bf97027" src="https://github.com/user-attachments/assets/af156fea-375b-4ace-9676-64af53dbe8ad" />
Code for estimating country-level musical opinion signals and comparing
multidimensional DeGroot, Friedkin--Johnsen (FJ), and state-dependent DeGroot
(SD-DeGroot) models.

This repository intentionally contains **source code only**. Datasets, model
weights, embeddings, intermediate files, figures, and experiment outputs are
excluded.

## Current experiment

- Period: 1875--1940
- Agents: Russia, France, Germany, Austria, and Others
- Representation: ChordBERT work embeddings
- State space: PC1--PC23 (more than 90% cumulative explained variance)
- Signal extraction: causal Kalman filtering
- Training transitions: 1875--1925
- Recursive forecasts: 1930, 1935, and 1940
- Models:
  - multidimensional DeGroot;
  - multidimensional FJ;
  - scale-normalized SD-DeGroot.

The PCA-weighted specifications use normalized explained-variance ratios in the
estimation loss. Weighting does not alter the transition equation or introduce
cross-component interactions.

## Repository layout

```text
src/                         ChordBERT, PCA, Kalman, and Python dynamics code
opinion_dynamics/
  degroot_fj/                Variance-weighted DeGroot and FJ
  fj/                        Independent multidimensional FJ implementation
  sd_degroot/                Scale-normalized SD-DeGroot and figure generators
  reproducibility/           Wolfram/Node experiment runners
mathematica/                 Original reference notebooks
chord_embedding/chordbert/   Chord parsing and model inference utilities
midi_experiment/src/         Optional MIDI/BiLSTM supplement
```

## Data and model setup

The main data source is the
[Cross-era Dataset](https://www.audiolabs-erlangen.de/resources/MIR/cross-era).
Composition years can be obtained from
[cross-era-classical-work-years](https://huggingface.co/datasets/hreyulog/cross-era-classical-work-years).

Large assets are not versioned. Place or generate them under the paths expected
by the scripts:

```text
data/                         Raw and processed datasets
chord_embedding/dataset/      Generated chord-sequence splits
chord_embedding/chordbert/    Downloaded or trained model directories
output/                       Generated embeddings, signals, metrics, and figures
```

All these paths are ignored by Git.

## Installation

Python 3.10 or newer is recommended.

```bash
python -m venv .venv
pip install -r requirements.txt
```

The opinion-dynamics scripts additionally require Wolfram Language 13.3 or
newer. Figure SVG generators use Node.js 18 or newer and only built-in modules.

## Main Python pipeline

```bash
python src/preprocess.py
python src/build_chordbert_datasets.py
python src/evaluate_chordbert.py
python src/analyze_chordbert_pca_1875_1940.py
python src/evaluate_degroot_chordbert_window_minmax.py
python src/evaluate_fj_chordbert_window_minmax.py
```

Scripts resolve paths relative to the repository root and create `output/`
when needed.

## Current opinion-dynamics models

Run from the repository root:

```powershell
wolframscript -file opinion_dynamics/degroot_fj/VarianceWeighted_DeGroot_FJ_PC1_PC23.wls
wolframscript -file opinion_dynamics/sd_degroot/ScaleNormalized_SDDeGroot_Equal_and_PCAWeighted_PC1_PC23.wls
wolframscript -file opinion_dynamics/sd_degroot/Export_ActualState_W_Diagnostics.wls
```

The scale-normalized SD-DeGroot model uses

```text
U_ij(t) = log(B_j) - h_i d_ij(t)/tau_d + Gamma_j q_j(t)/tau_q,
```

where `B` is a positive simplex, `d_ij` is RMS stylistic distance, `q_j` is
RMS distance from the cross-country centroid, and `tau_d` and `tau_q` are
training-period medians. It uses no min--max scaling, clipping, or held-out
information.

## Reproducibility and leakage control

- Kalman filtering is causal.
- Training uses states through 1925 only.
- Forecasts for 1930--1940 are recursive.
- PCA variance weights affect estimation losses only.
- Actual-state influence matrices for 1930 and 1935 are diagnostic and are
  never used to produce forecasts or evaluation metrics.

## Optional MIDI supplement

See [midi_experiment/README.md](midi_experiment/README.md). Its external dataset
and model weights are intentionally excluded.

## Citation
```bibtex
@misc{he2026modelingstylisticcoevolutionsymbolic,
      title={Modeling Stylistic Co-evolution in Symbolic Music Heritage Collections}, 
      author={Yulong He and Ivan Smirnov and Yanming Li},
      year={2026},
      eprint={2607.23957},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2607.23957}, 
}