Decode the Beat: Spotify Genre Classification

Advanced Multi-Class Classification Workflow using Class-Balanced Voting Ensemble

---

Objective

Build a robust machine learning classification model to accurately predict the track_genre of Spotify songs based on distinct audio features and metadata (without AutoML tools, external APIs, or pre-trained models).

DataSprint - Data Science Club, NIST University — Kaggle Hackathon Link

Attributes	Description
track_genre	Target categorical variable (e.g., pop, rock, classical)
danceability, energy, tempo	Core acoustic and rhythmic features
acousticness, liveness, valence	Tonal and mood indicators
duration_ms, time_signature	Track structural and length data

---

Dataset

Metric	Value
Training tracks	84,800
Test tracks	34,200 (unlabeled flat test.csv)
Format	.csv (Tabular numeric and categorical data)

Features Overview: The datasets (train.csv, test.csv) contain numerous acoustic vectors representing the track's sound profile alongside entity metadata (e.g., artists). The target track_genre is only present in the train split.

---

Methodology

Step 1 - Preprocessing

• Missing numeric and categorical data were handled independently utilizing median/mode-based imputation.
• Executed Frequency Encoding explicitly targeting the artists feature.
• Normalization: Applied StandardScaler to ensure features contributed uniformly across internal distance metrics and split criteria.

Step 2 - Feature Engineering: Acoustic Intersections

Derived meaningful new dimensions from raw inputs to amplify predictive signals.

• Synthesized interaction terms representing acoustic dynamics (e.g., energy_x_danceability).
• Identified mood_divergence metrics through overlapping feature spaces.

Enhancing the base dimensional space allowed tree-based algorithms to discover richer splits early in their decision paths.

Step 3 - Model Architecture: Voting Ensemble Setup

Tackled complex, overlapping music genre clusters by developing a Class-Balanced Voting Ensemble, leveraging complementary tree structures.

LightGBM (Gradient Boosting) -> efficiently fits deep non-linear gradients to accurately identify dense clusters. Extra Trees (Extremely Randomized Trees) -> mitigates overfitting by introducing extreme variance during feature partitioning.

Ensemble voting captures both deep contextual gradient boundaries and generalized random boundaries for stable, generalized performance.

Step 4 - Validation: Out-of-Fold (OOF)

Adopted a robust Cross-Validation architecture avoiding traditional validation holdouts. Stratified Out-of-Fold logic evaluated metrics over the entire dataset, serving as a highly reliable generalized estimator for the unseen 34,200 test tracks.

---

Results

Evaluated across the entire dataset via Out-of-Fold (OOF) validation strategy

Metric	Score
Accuracy Assessment	0.3321
Precision Aggregate	0.3291
F1 Score Evaluation	0.3276

---

Repository Structure

Spotify Track Genre Classification
├── NOTICE
├── LICENSE 
├── README.md                               -> High-level summary (This file)
├── data/                                   -> Provided datasets (`train.csv`, `test.csv`)
├── report.txt                              -> Detailed approach and methodology report
├── submission.csv                          -> Final Kaggle prediction submission (34,200 rows)
└── track_genre_classification.ipynb        -> Full Jupyter Notebook - preprocessing, training, evaluation

---