Spotify Genre Segmentation & Recommendation System - Minor Project

This is my minor project for the AI & ML course. I built a system that groups Spotify tracks into clusters based on their audio features (acousticness, danceability, energy, etc.) and uses these clusters to make similar song recommendations.

---

🛠️ Data Cleaning & Preprocessing

To get clean clusters, I preprocessed the data:

1. Removed Index Column: Removed the duplicate row indices.
2. Missing Metadata: Dropped any rows where critical fields like track name or artist name were empty.
3. Duplicates: Spotify datasets often have duplicate tracks if they are on multiple playlists. I dropped duplicate tracks based on track_id (removed 24,259 tracks).

• Final Dataset: Kept 89,740 unique tracks (78.72% of the original data).

---

📈 Key Insights from Data Analysis (EDA)

I generated several exploratory plots (saved in the plots/ folder):

• audio_features_distributions.png: Shows the distribution of features like danceability, energy, valence, and tempo.
• audio_features_correlation.png: Correlation matrix of acoustic properties. Shows that energy is strongly positively correlated with loudness (+0.82) and strongly negatively correlated with acousticness (-0.73).
• top_genres.png: Displays track counts for the most popular genres.
• danceability_energy_by_genre.png: Boxplots showing how acoustic profiles (energy and danceability) vary significantly between specific genres.

---

🧼 Clustering (K-Means)

To group the songs, I scaled the audio features using StandardScaler and ran K-Means:

1. Elbow Method: Plotted WCSS (Within-Cluster Sum of Squares) against $K$ (range 2 to 12) using a representative sample of 15,000 tracks. The resulting curve is saved as plots/elbow_curve.png.
2. Final Clusters: Based on the elbow curve, I selected K=10 clusters. The 2D PCA projection of the clusters is saved as plots/cluster_visualization.png.

Cluster Profiles & Characteristics:

• Cluster 0: High energy, low acousticness (contains death-metal, black-metal, hardstyle).
• Cluster 1: Acoustic, slow, romantic songs (contains romance, honky-tonk, tango).
• Cluster 8: Very quiet, slow ambient music (contains new-age, sleep, ambient).
• Cluster 4: Upbeat, high-energy pop and dance music (contains dance, k-pop, turkish).

---

📻 Similarity-Based Recommendation Engine

I built a similarity-based recommendation engine (recommend_spotify.py) using Cosine Similarity:

1. It searches for a user-inputted song name.
2. It finds which cluster that song belongs to.
3. It filters the dataset to only search inside that same cluster (this keeps recommendations accurate and fast).
4. It computes the cosine similarity between the scaled audio features of your song and all other songs in the cluster.
5. It returns the top 5 closest matches.

Example Recommendation Run:

• Input: "Hold On" by Chord Overstreet (Acoustic, Cluster 1)
• Recommendations:
  1. Hold On by Chord Overstreet (Similarity: 0.9997)
  2. ?? by Jay Chou (Similarity: 0.9831)
  3. Out Of Time by Lund (Similarity: 0.9726)
  4. La Llamada by Leiva (Similarity: 0.9707)
  5. Acredita by Novo Som (Similarity: 0.9677)

---

📁 Project Structure

spotify-genre-segmentation/
├── preprocess_spotify.py    # Data cleaning and feature scaling
├── visualize_spotify.py     # EDA and audio feature correlation
├── cluster_spotify.py       # K-Means clustering and evaluation
├── recommend_spotify.py     # Cosine similarity-based recommendation engine
├── requirements.txt         # Dependencies
├── README.md                # This file
├── LICENSE                  # MIT License
├── spotify_kmeans_model.pkl # Serialized K-Means model
├── spotify_scaler.pkl       # Serialized feature scaler
└── plots/                   # Saved clustering and EDA plots

---

🚀 How to Run

1. Install requirements:

   pip install -r requirements.txt

2. Put spotify_dataset.csv in the same directory.
3. Run preprocessing:

   python preprocess_spotify.py

4. Run plots:

   python visualize_spotify.py

5. Run clustering:

   python cluster_spotify.py

6. Get recommendations:

   python recommend_spotify.py

---

📜 License

This project is open-source under the MIT License. See LICENSE for details.