Project
Node Classification
(1) Homophily and Heterophily Analysis of Node Classification
- Find and Download Datasets for Node Classification dataset link
- Calculate Homophily and Heterophily of different networks
- Code GNN-based models and obtain node classification performance
- Analyze the relationship between the homophily of a network/dataset and its corresponding node classification performance
(2) Oversmoothing Analysis of Node Classification
- Find and Download Datasets for Node Classification dataset link
- Code GNN-based models and obtain node classification performance
- Change the Message-passing layer and obtain performance for each layer
- Analyze the relationship between the number of message-passing layers and the node classification performance
(3) Node Classification on Heterogeneous Graphs
- Find and Download Heterogeneous Graph Datasets for Node Classification dataset link
- Code Heterogeneous GNN-based models and obtain node classification performance
Link Prediction
(1) Homophily and Heterophily Analysis of Link Prediction
- Find and Download Datasets for Link Prediction dataset link
- Calculate Homophily and Heterophily of different networks
- Code GNN-based models and obtain link prediction performance
- Analyze the relationship between the homophily of a network/dataset and its corresponding link prediction performance
(2) Degree Analysis of Link Prediction
- Find and Download Datasets for Link Prediction dataset link
- Calculate Degree of different nodes within one network/dataset
- Code GNN-based model and obtain link prediction performance for a specific network
- Calculate the link prediction performance for each node and group nodes according to their degree in different groups
- Analyze the relationship between the average group link prediction performance and their average group degree
(3) Implement LightGCN-based Recommender System and Compare its performance on users with different degrees/activity
- Find and Download Datasets for the Recommender System dataset link
- Code LightGCN as your GNN-based recommender system and obtain recommendation performance for a specific network
- Calculate the recommendation performance for each user and group them based on user degree
- Analyze the relationship between the average group recommendation performance and their average group degree
- Read materials related to cold-start problems.
(4) Implement LightGCN-based Recommender System and Compare its performance with and without text information
- Find and Download Datasets for the Recommender System dataset link
- Code LightGCN as your GNN-based recommender system and obtain recommendation performance for a specific network under two scenarios: with text and without text
- Calculate the recommendation performance for each user and group them based on user degree
- Analyze the relationship between the average group recommendation performance and their average group degree
- Read materials related to cold-start problems.
Edge Classification
(4) Compare Different GNN Models and MLP models on Edge Classification
- Find and Download Datasets for the Recommender System dataset link
- Process the review information for each customer-product interaction
- Code GNN to implement edge classification
- Calculate the performance and analyze the relationship between the edge classification performance with degree of ending points
Graph Classification
(5) Investigate the Imbalance Issue on Graph Classification
- Find and Download Datasets for the Graph Classification dataset link
- Calculate the number of graphs in different classes
- Code GNN to implement graph classification
- Change the number of training/validation/testing graphs to create balance and imbalance situations
- Train the GNN model, obtain graph classification performance on different imbalance ratios, and analyze their relationship.
(6) Investigate whether incorporate 3D coordinates would help Graph Classification
- Find and Download Datasets for Molecule Classification dataset link
- Code GNN to implement graph classification for 3D graphs
- Compare Graph Classification performance with and without 3D coordinates.