CitationNetworkLoader<T>
Models & Types · AiDotNet.Data.Graph
Loads citation network datasets (Cora, CiteSeer, PubMed) for node classification.
Inheritance: DataLoaderBase<T> → GraphDataLoaderBase<T> → CitationNetworkLoader<T>
For Beginners
Citation networks are graphs of research papers.
Structure:
- Nodes: Research papers
- Edges: Citations (Paper A cites Paper B)
- Node Features: Bag-of-words representation of paper abstracts
- Labels: Research topic/category
Datasets:
Cora:
- 2,708 papers
- 5,429 citations
- 1,433 features (unique words)
- 7 classes (topics): Case_Based, Genetic_Algorithms, Neural_Networks,
Probabilistic_Methods, Reinforcement_Learning, Rule_Learning, Theory
- Task: Classify papers by topic
CiteSeer:
- 3,312 papers
- 4,732 citations
- 3,703 features
- 6 classes: Agents, AI, DB, IR, ML, HCI
PubMed:
- 19,717 papers (about diabetes)
- 44,338 citations
- 500 features
- 3 classes: Diabetes Mellitus Type 1, Type 2, Experimental
Key Property: Homophily Papers tend to cite papers on similar topics. This makes GNNs effective:
- If neighbors are similar topics, aggregate their features
- GNN learns to propagate topic information through citation network
- Even unlabeled papers can be classified based on what they cite
How It Works
Citation networks are classic benchmarks for graph neural networks. Each dataset represents academic papers as nodes and citations as edges, with the task being to classify papers into research topics.
Constructors
| Constructor | Summary |
|---|---|
CitationNetworkLoader(CitationDataset<T> dataset, string dataPath, bool autoDownload) |
Initializes a new instance of the CitationNetworkLoader class. |
Properties
| Property | Summary |
|---|---|
int NumClasses { get; } |
|
string Description { get; } |
|
string Name { get; } |
Methods
| Method | Summary |
|---|---|
GraphClassificationTask<T> CreateGraphClassificationTask(double trainRatio, double valRatio, int? seed) |
|
LinkPredictionTask<T> CreateLinkPredictionTask(double trainRatio, double negativeRatio, int? seed) |