Add V31_FEATURE_ENGINEERING_NOTES.md
Browse files
V31_FEATURE_ENGINEERING_NOTES.md
ADDED
|
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# V31 Feature Engineering Notes
|
| 2 |
+
|
| 3 |
+
## Why we did V31
|
| 4 |
+
V16 is stable and strong, but it may saturate. V31 adds richer signal descriptors to capture information that plain MFCC/chroma stats can miss.
|
| 5 |
+
|
| 6 |
+
Goal:
|
| 7 |
+
- keep V16 baseline safety,
|
| 8 |
+
- add meaningful audio dynamics/rhythm/texture features,
|
| 9 |
+
- compare objectively against V16 before upload.
|
| 10 |
+
|
| 11 |
+
## What was added in V31
|
| 12 |
+
1. **MFCC Delta + Delta-Delta**
|
| 13 |
+
- Captures temporal changes in timbre.
|
| 14 |
+
- Useful when genres differ by evolving patterns, not just static spectral shape.
|
| 15 |
+
|
| 16 |
+
2. **HPSS (Harmonic-Percussive Source Separation)**
|
| 17 |
+
- Splits audio into harmonic and percussive components.
|
| 18 |
+
- Helps distinguish melody-heavy vs beat-heavy genre characteristics.
|
| 19 |
+
|
| 20 |
+
3. **Tempo / Beat Strength Features**
|
| 21 |
+
- Adds rhythmic cues (tempo, onset strength mean/std).
|
| 22 |
+
- Useful for genre rhythm identity.
|
| 23 |
+
|
| 24 |
+
## Where implemented
|
| 25 |
+
- Notebook: `notebooks/DL_21f2000735_FINAL_UPLOAD_V31.ipynb`
|
| 26 |
+
- Demo feature output script: `src/v31_feature_demo_output.py`
|
| 27 |
+
- Demo output file: `reports/v31_demo_feature_output.json`
|
| 28 |
+
|
| 29 |
+
## How to use it in decisions
|
| 30 |
+
- Run V31 compare block (V16-base vs V31-fusion).
|
| 31 |
+
- Check:
|
| 32 |
+
- `v16_base_macro_f1`
|
| 33 |
+
- `v31_fusion_macro_f1`
|
| 34 |
+
- `delta_v31_minus_v16`
|
| 35 |
+
- Upload V31 only if delta is non-negative and run is stable.
|
| 36 |
+
|
| 37 |
+
## How to learn more
|
| 38 |
+
1. **Librosa docs**
|
| 39 |
+
- MFCC / delta / HPSS / tempo APIs
|
| 40 |
+
2. **Audio feature engineering guides**
|
| 41 |
+
- spectral contrast, rhythm features, pooling strategies
|
| 42 |
+
3. **Ablation practice**
|
| 43 |
+
- turn one feature family on/off and record F1 delta
|
| 44 |
+
4. **Error analysis loop**
|
| 45 |
+
- inspect confusion matrix and add features targeting weak class pairs
|
| 46 |
+
|
| 47 |
+
## Practical next learning experiments
|
| 48 |
+
1. HPSS-only ablation vs base
|
| 49 |
+
2. Delta-only ablation vs base
|
| 50 |
+
3. Tempo-only ablation vs base
|
| 51 |
+
4. Combined (V31) vs V16 baseline
|
| 52 |
+
5. Class-wise F1 comparison for weak genres
|
| 53 |
+
|
| 54 |
+
This note exists to keep V31 explainable in viva and reproducible in future runs.
|