A comprehensive music recommendation system implementing multiple approaches from baseline methods to advanced deep learning architectures.

This project implements a complete music recommendation pipeline with four different approaches, each building upon the previous one to create increasingly sophisticated recommendation systems.

Music Recommendation/
├── 0.data_exploration.py          # Data analysis and exploration
├── 1.KNN_based.py                 # KNN-based recommendations
├── 2.RandomForest_based.py        # RandomForest ranking model
├── 3.SASRec_sequential.py         # SASRec sequential recommendations
├── 4.Two_tower.py                 # Two-tower architecture with FAISS
├── requirements.txt               # Python dependencies
└── README.md                      

• Python 3.8+
• CUDA-compatible GPU (optional, for GPU acceleration)
• 8GB+ RAM recommended

1. Clone or download the project

2. Install dependencies:

   pip install -r requirements.txt

3. Run any stage:

   python 0.data_exploration.py
   python 1.KNN_based.py
   python 2.RandomForest_based.py
   python 3.SASRec_sequential.py
   python 4.Two_tower.py

Purpose: Understand datasets and log insights

Features:

• Dataset statistics and distributions
• Audio feature analysis
• User behavior patterns
• Artist popularity analysis
• Weights & Biases integration

Output:

• Comprehensive data insights
• W&B logged metrics
• Dataset summaries

Purpose: Baseline recommendation methods

Methods:

• Popularity-based: Most popular artists
• Content-based kNN: Audio feature similarity
• Hybrid approach: Combines both methods

Features:

• Cosine similarity for audio features
• StandardScaler for feature normalization
• Comprehensive evaluation metrics

Output:

• Hit Rate and NDCG metrics
• Top popular artists
• Similar song recommendations

Purpose: Learning-to-Rank approach

Features:

• Feature Engineering: User behavior + audio features
• Random Forest: 100 estimators for ranking
• Derived Features: Weight ratios, diversity metrics
• Feature Importance: Identifies key factors

Key Features:

• User activity patterns
• Artist popularity metrics
• Interaction strength analysis
• Relative activity measures

Output:

• RMSE, MAE, Correlation metrics
• Feature importance rankings
• Personalized recommendations

Purpose: Sequential recommendation modeling

Architecture:

• SASRec Model: Self-Attentive Sequential Recommendation
• Transformer: 4 heads, 2 layers, 128 hidden size
• Embeddings: Item + Position embeddings
• Sequence Length: Up to 50 items per sequence

Features:

• GPU acceleration with mixed precision
• Sequence padding and truncation
• Cross-entropy loss with padding ignore
• Comprehensive evaluation (Hit Rate@K, NDCG@K)

Output:

• Sequential recommendations
• Training/validation loss curves
• Ranking metrics at multiple K values

Purpose: Advanced neural collaborative filtering

Architecture:

• User Tower: Neural network for user embeddings
• Item Tower: Neural network for item embeddings
• FAISS Retrieval: Fast similarity search
• Cold-Start Handling: New users/items support

Features:

• GPU Optimization: Mixed precision training
• Feature Engineering: Log-scaled heavy-tailed features
• Comprehensive Evaluation: Precision@K, Recall@K, NDCG@K
• Cold-Start Solutions: Average embeddings fallback
• FAISS Index: IVF-based fast retrieval

Output:

• High-quality embeddings
• Fast retrieval recommendations
• Comprehensive evaluation metrics
• Cold-start recommendations

• CUDA Support: Automatic GPU detection
• Mixed Precision: 16-bit training for speed
• Memory Optimization: Efficient tensor operations
• Batch Processing: Optimized batch sizes

• Precision@K: Accuracy of top-K recommendations
• Recall@K: Coverage of relevant items
• NDCG@K: Ranking quality measure
• Hit Rate@K: Success rate of recommendations

• New Users: Average user embedding fallback
• New Items: Average item embedding fallback
• Feature Engineering: Robust feature creation
• Graceful Degradation: System continues working

Below data recorded for less epochs(~10)

Performance may vary based on hardware and data

• Core: pandas, numpy, scikit-learn
• Deep Learning: torch, transformers
• Retrieval: faiss-cpu
• Tracking: wandb
• Visualization: matplotlib, seaborn, plotly

# Run data exploration
python 0.data_exploration.py

# Run KNN recommendations
python 1.KNN_based.py

# Run RandomForest ranking
python 2.RandomForest_based.py

# Run SASRec sequential
python 3.SASRec_sequential.py

# Run Two-Tower architecture
python 4.Two_tower.py

• Automatic Logging: Metrics, losses, and visualizations
• Experiment Tracking: Compare different runs
• Hyperparameter Tuning: Track parameter effects
• Model Comparison: Side-by-side performance analysis

• Progress Bars: Training progress visualization
• Console Output: Real-time metrics and status
• Error Handling: Graceful failure management

• SASRec: Self-Attentive Sequential Recommendation
• Two-Tower: Neural Collaborative Filtering
• FAISS: Facebook AI Similarity Search
• Last.fm Dataset: HetRec 2011 Challenge
• Spotify API: Audio Features Documentation