Spotify-FMA-TracksPopularity-Prediction

Project title and overview

Initial Project Proposal

This project predicts Hip-Hop track popularity using two complementary datasets:

• Spotify: Metadata-based prediction (13 features)
• FMA: Rich audio features (524 features) + CNN on spectrograms Target: 3-class popularity classification (Low / Medium / High)

Source of Datasets

1. Spotify Tracks Dataset

• Source: Kaggle - Spotify Tracks Dataset
• License: CC0: Public Domain
• Size: 232+ tracks
• Features: 18 attributes including:
  • popularity: Target variable (0-100)
  • danceability: How suitable for dancing (0.0-1.0)
  • energy: Intensity measure (0.0-1.0)
  • acousticness: Acoustic vs electric (0.0-1.0)
  • valence: Musical positivity (0.0-1.0)
  • tempo: Beats per minute (BPM)
  • ....
• Example:

  Artist: Drake
  Track: God's Plan
  Danceability: 0.754 | Energy: 0.449 | Valence: 0.357
  Duration: 3:18 | Tempo: 77 BPM
  Artist Tracks: 156 | Avg Popularity: 84.2

  Artist: Unknown Underground
  Track: Late Night Session
  Danceability: 0.623 | Energy: 0.512 | Valence: 0.441
  Duration: 2:47 | Tempo: 92 BPM
  Artist Tracks: 1 | Avg Popularity: 0.0

2. FMA (Free Music Archive)

• Source: FMA Dataset on GitHub FMA GitHub
• License: Creative Commons Attribution 4.0 International (CC BY 4.0)
• Subset: Hip-Hop tracks (8,389 tracks)
• Features: After preprocessed 525 features (518 audio + 7 metadata)
• Example:

    Artist: Doomtree
    Album: No Kings
    Track: The Grand Experiment
    Duration: 258s | Bit Rate: 320 kbps
    Comments: Album: 10, Artist: 20, Track: 10
    .....
  

Installing the project

1. Clone the repository

   git clone <repo-url>
   cd <repo-folder>

2. Install dependencies and run the full pipeline The project includes a Makefile that automates:

   make all

   View Available Commands Type

   make help

This will:

1. Install Python dependencies
2. Download Spotify dataset from Kaggle
3. Preprocess Spotify data (filter Hip-Hop, engineer features, split, scale)
4. Download FMA metadata and audio
5. Preprocess FMA metadata (extract 525 features)
6. Convert MP3 files → spectrograms for CNN

Optiona: Train Models (Pre-trained models available) make train # Train all models (asks for consent) make train_spotify # Train Spotify models only make train_fma # Train FMA models only make train_cnn # Train CNN (takes 0.5-1+ hours)

Optional: Evaluate Models make evaluate # Evaluate all models (asks for consent) make eval_spotify # Evaluate Spotify models make eval_fma # Evaluate FMA models make eval_cnn # Evaluate CNN

Data Preparation Pipeline

The entire data preparation workflow is fully automated using the project’s Makefile.Running the command below downloads datasets, preprocesses them, and generates consistent outputs for all team members. The project includes a Makefile that automates:

make all

View Available Commands Type

make help

This will:

1. Install Python dependencies
2. Download Spotify dataset from Kaggle
3. Preprocess Spotify data (filter Hip-Hop, engineer features, split, scale)
4. Download FMA metadata and audio
5. Preprocess FMA metadata (extract 525 features)
6. Convert MP3 files → spectrograms for CNN

Optiona: Train Models (Pre-trained models available) make train # Train all models (asks for consent) make train_spotify # Train Spotify models only make train_fma # Train FMA models only make train_cnn # Train CNN (takes 0.5-1+ hours)

Optional: Evaluate Models make evaluate # Evaluate all models (asks for consent) make eval_spotify # Evaluate Spotify models make eval_fma # Evaluate FMA models make eval_cnn # Evaluate CNN

FMA Preprocessing Modes

FMA has two preprocessing modes:

1. Full Mode (default) - 8,389 tracks

   make fma_metadata_full

   • Uses all Hip-Hop tracks
   • For baseline ML model training
   • Output: X_train.csv, y_train.csv, etc.

2. Fair Mode - 997 tracks (CNN comparison)

   make fma_metadata_fair

   • Only tracks with spectrograms
   • For fair CNN vs ML comparison
   • Output: X_train_fair.csv, y_train_fair.csv, spectrogram_labels.csv

Evaluation Spotify Baseline models Results Summary:

              Model Data Type  CV Accuracy  CV Accuracy Std    CV F1  CV F1 Std  Test Accuracy  Test Precision  Test Recall  Test F1
       RandomForest       raw     0.555272         0.010682 0.534611   0.010692       0.572889        0.560561     0.560372 0.550824
Logistic Regression    scaled     0.560787         0.008327 0.544361   0.008297       0.580958        0.564873     0.570945 0.561667
                KNN    scaled     0.484667         0.015347 0.463354   0.016885       0.485207        0.469237     0.475988 0.466237
         GaussianNB       raw     0.522728         0.007008 0.494348   0.005311       0.532544        0.508015     0.518863 0.504699



Evaluation FMA baseline models Results Summary:

              Model Data Type  CV Accuracy  CV Accuracy Std    CV F1  CV F1 Std  Test Accuracy  Test Precision  Test Recall  Test F1
       RandomForest       raw     0.416903         0.030448 0.416804   0.030714           0.32        0.304762     0.305271 0.304974
         GaussianNB       raw     0.416977         0.014716 0.404621   0.016415           0.26        0.213238     0.246752 0.203595
                KNN    scaled     0.364538         0.033346 0.348658   0.033960           0.38        0.397974     0.396068 0.363698
Logistic Regression    scaled     0.396903         0.024682 0.396516   0.024589           0.36        0.363585     0.367464 0.357262

======================================================================
CNN MODEL EVALUATION - TEST SET RESULTS
======================================================================

Model: cnn_fma_production
Test samples: 100

Metric                         Value
-------------------------------------
Accuracy                      0.3800
Precision (macro)             0.2608
Precision (weighted)          0.2931
Recall (macro)                0.3333
Recall (weighted)             0.3800
F1-Score (macro)              0.2832
F1-Score (weighted)           0.3204


PER-CLASS PERFORMANCE ANALYSIS
======================================================================

Low:
  True samples:     39
  Predicted as Low: 68
  Correctly predicted: 26
  Precision:        0.3824
  Recall:           0.6667
  F1-Score:         0.4860

Medium:
  True samples:     36
  Predicted as Medium: 30
  Correctly predicted: 12
  Precision:        0.4000
  Recall:           0.3333
  F1-Score:         0.3636

High:
  True samples:     25
  Predicted as High: 2
  Correctly predicted: 0
  Precision:        0.0000
  Recall:           0.0000
  F1-Score:         0.0000


======================================================================
PERFORMANCE COMPARISON: CNN vs Baseline Models (FMA - 997 tracks)
======================================================================
              Model  Accuracy  Precision  Recall  F1-Score
               k-NN      0.40   0.397596    0.40  0.364405
  CNN (Spectrogram)      0.38   0.293118    0.38  0.320442
      Random Forest      0.32   0.319143    0.32  0.319524
Logistic Regression      0.38   0.151224    0.38  0.216350
        Naive Bayes      0.26   0.215033    0.26  0.210764
======================================================================

Spotify-FMA-TracksPopularity-Prediction