DEAM_Preprocessing_Guide.md

🧹 DEAM Dataset Preprocessing Task Guide

This guide outlines the preprocessing steps needed to prepare the DEAM dataset for training a mood classification neural network. It is optimized to be used with AI-assisted coding tools like Cursor.

---

🎯 Objective

Create a training-ready dataset by merging DEAM's audio features with averaged valence/arousal annotations, and optionally generate discrete mood labels for classification.

---

✅ Task Breakdown

1. Load Song-Level Audio Features

• Source: features/audio_features/
• Action:
  • Load all relevant CSVs.
  • Merge them into a single DataFrame using song_id or consistent indexing.
  • Drop any duplicate or NaN rows.

---

2. Load Averaged Valence & Arousal Labels

• Source: annotations/averaged_annotations/song_level/static_annotations_averaged_songs_*.csv
• Action:
  • Extract song_id, valence_mean, and arousal_mean columns.
  • This will serve as your target label(s).

---

3. Merge Features with Labels

• Join the features DataFrame with the labels using song_id or filename.
• Drop rows with missing or misaligned values.

---

4. (Optional) Generate Discrete Mood Labels

• Create a new column mood_class using this logic:

def mood_label(valence, arousal):
    if valence >= 0.5 and arousal >= 0.5:
        return "Energetic"
    elif valence >= 0.5 and arousal < 0.5:
        return "Chill"
    elif valence < 0.5 and arousal >= 0.5:
        return "Tense"
    else:
        return "Melancholic"

---

5. Save the Final Processed Dataset

• Output file: processed_dataset.csv
• Columns should include:
  • Feature columns (e.g. tempo, mfcc_1, rms, ...)
  • valence, arousal
  • (Optional) mood_class

---

📦 Final Output Format

tempo	mfcc_1	...	valence	arousal	mood_class
120.0	0.123	...	0.71	0.62	Energetic

---

💡 Notes

• Normalize or standardize features if needed during model training.
• Save as .csv or .parquet for fast loading later.
• Ensure reproducibility by setting a random seed for any data splits.

---