Graph Model Optimization for Music Recommendation

Overview

This project explores the optimization of graph-based models for music recommendation systems. The aim is to enhance recommendation accuracy by refining the algorithmic framework to better capture user-music interactions. The implementation leverages the NetworkX and StellarGraph libraries to construct and analyze heterogeneous graphs representing user and music data.

Objectives

• Construct a music recommendation system using heterogeneous graphs.
• Address challenges in existing graph-based techniques like efficiency and scalability.
• Optimize graph models to improve recommendation accuracy and user satisfaction.

Research Questions

1. How can heterogeneous graphs be used to model relationships in music data?
2. What benefits do StellarGraph and NetworkX bring to music recommendation systems?
3. How do these optimizations influence user satisfaction and engagement?

Methodology

Libraries and Tools

• NetworkX: Used for basic graph operations and prototyping.
• StellarGraph: Handles large-scale graph data suitable for advanced machine learning tasks.
• HinSAGE Algorithm: Extends GraphSAGE to work with heterogeneous graph data, making it suitable for complex datasets.

Data

The project uses the 30Music dataset, which includes user playlists, listening sessions, positive ratings, and more. This dataset is crucial for predicting user preferences and making playlist recommendations. https://recsys.deib.polimi.it/datasets/

Graph Construction

1. Neighborhood Sampling: For each node, a fixed-size neighborhood is sampled, involving neighbors from different node types.
2. Aggregation: Features of the sampled neighbors are aggregated. In heterogeneous graphs, this can be type-specific.
3. Concatenation and Final Prediction: Embeddings of node pairs are concatenated and passed through a final estimator layer to produce the predicted score.

Evaluation Metrics

• Mean Squared Error (MSE)
• Mean Absolute Error (MAE)
• Root Mean Square Error (RMSE)

Results

• The model significantly outperformed the baseline in terms of MAE and RMSE.
• Visualization techniques like histograms and scatter plots were used to understand data patterns and relationships.

Usage

Requirements

• Python 3.x
• NetworkX
• StellarGraph
• Other dependencies as specified in the Jupyter notebook

Running the Notebook

1. Clone the repository.
2. Install the required dependencies using pip install -r requirements.txt.
3. Open and run the Music_Recommendation.ipynb notebook.

pip install -r requirements.txt
jupyter notebook Music_Recommendation.ipynb

Code Structure

• Data Preparation: Loading and preprocessing the 30Music dataset.
• Graph Construction: Building heterogeneous graphs using NetworkX and StellarGraph.
• Model Training: Implementing the HinSAGE algorithm to train the recommendation model.
• Evaluation: Assessing model performance using MAE and RMSE.

Conclusion

This project demonstrates the effectiveness of using optimized graph models for music recommendation systems. By leveraging NetworkX and StellarGraph, we can handle complex user-music interactions and provide personalized recommendations with improved accuracy.