This project tests whether East Asian influence in Western classical piano — the gamelan- and pentatonic-derived writing found in Debussy and Ravel — can be detected from symbolic musical features alone, and which features most reliably predict it. The goal is to replace case-by-case subjective attribution with a systematic, reproducible method.
Corpus & labels
- 165 four-bar segments across 12 solo-piano works (piano-only, to control for instrumentation).
- Influenced: Debussy (“Pagodes,” “Et la lune descend,” “Cloches,” “Canope”) and Ravel (“Laideronnette”). Controls: Bach, Beethoven, Mozart, Brahms.
- Influence labels and per-segment feature scores assigned from documented musicological evidence.
Features
Nine symbolic features, each hand-scored on a 0–2 scale: pentatonicism, parallel motion, density, rhythmic regularity, syncopation, melodic intervals, register usage, articulation, and dynamics. Two are theory-motivated markers (pentatonicism, parallel motion); the rest are exploratory controls to probe model sensitivity.
Model & evaluation
- Random Forest — chosen for interpretability and data-efficiency on a small, mixed-scale tabular set, where linear models assume separability and deep nets overfit at n=165.
- Group-aware throughout to prevent leakage between segments of the same piece: a frozen ≈65/35 grouped hold-out (GroupShuffleSplit) plus grouped 5-fold StratifiedGroupKFold for descriptive analysis.
- Hyperparameters tuned by RandomizedSearchCV on positive-class F1; final model
n_estimators=200, max_depth=10, fixed seed.
Results
- Hold-out: accuracy 0.656, balanced accuracy 0.658, AUROC 0.809, AUPRC 0.875 for the influenced class.
- Precision 0.750 / recall 0.649 — the model is conservative: it ranks segments well but under-calls borderline influence.
- Pentatonicism dominates — ~32% of all depth-≤2 tree splits and the largest hold-out permutation hit (ΔAUPRC ≈ −0.20). It acts as a gate: sustained pentatonic writing drives an “influenced” call; without it, noisier mid-tier cues (syncopation, density, parallel motion, articulation) decide.
- In practice: clean rejection of canonical Western works, oscillation on “Pagodes” as textures shift, and clustered false positives on short controls where mid-tier cues align without pentatonic support.
Takeaways & stack
The model confirms pentatonicism as the central marker (matching the literature) and surfaces syncopation and density as unexpected conditional amplifiers — while exposing how coarse symbolic coding flattens timbre, pedaling, and resonance. Built with Python: scikit-learn (Random Forest), music21 for symbolic parsing, pandas, NumPy, SciPy, and matplotlib. Fully reproducible — fixed seed, frozen split, and committed predictions and model.