bhavyasankhla / G-PLAY

Graph-based Playlist Continuation using Neural Collaborative Filtering

CMPT-733-Project

Project Proposal

Project Members

• Abhishek Deshpande, asd27
• Ayush Sharma, asa489
• Bhavya Sankhla, bsa89

Title

Automated Playlist Continuation using GNNs

Dataset

Spotify Million Playlist Challenge

The Spotify Million Playlist Dataset Challenge consists of a dataset and evaluation to enable research in music recommendations.

Technologies

• Python
• PyG
• PyTorch
• networkx
• DeepSnap
• graph-tool
• sklearn
• flask
• HTML/CSS
• ajax

Overview

Music recommendation systems have become an integral part of our daily lives, and collaborative filtering has been a popular technique used to predict user preferences. However, collaborative filtering suffers from several bottlenecks, including the lack of useful metadata, sparsity, and scalability issues. Graph-based collaborative filtering methods have gained popularity due to their ability to handle these bottlenecks and model complex relationships. In this project, we aim to use LightGCN, a state-of-the-art graph-based collaborative filtering method, to continue Spotify playlists by predicting which songs a user is likely to add to a playlist based on their existing preferences and playlist history. LightGCN leverages the graph structure to handle sparsity and scalability issues and incorporates neighbourhood information into nodes using an aggregation strategy. Additionally, it is capable of handling cold-start scenarios by generating embeddings for new users or items with no interactions by leveraging the embeddings of existing nodes in the graph. By improving the accuracy and scalability of music recommendation systems, we aim to provide a more personalized and satisfying music discovery experience for Spotify users, and ultimately contribute to the development of more effective and efficient recommendation systems in other domains.

How to Run Using Jupyter Notebook

1. Obtain the dataset from the Spotify Million Playlist Challenge on Aicrowd (try not to exceed 100 files).
2. Run the fixed_graph_gen.ipynb file and ensure that you update the file path within the notebook.
3. After running fixed_graph_gen.ipynb, a pickle file for the graph will be generated. You can then run the LightGCN_model.ipynb notebook to view the results within the notebook.

How to Run as a Web Application

1. Make sure you complete the above process of running the notebooks. This will generate a pickle file for our model.
2. Once the pickle file is generated, run the recommendation_server.py file. It will start a server and listen to requests on the "http://localhost:5000/recommend" endpoint.
3. Open the FrontEndWebPage.html file in a browser, enter an existing playlist name (e.g., "Spring16", "TOP", etc., which can be fetched from the dataset), and the server will respond with recommended songs and display them on the webpage.