DBpedia Entity Classification using GCNs and GraphSAGE

A multiclass text classification solution on the DBpedia-14 Ontology Classification Dataset using two advanced Graph Neural Network (GNN) architectures:

• GCN (Graph Convolutional Network) — Transductive setting
• GraphSAGE (Graph Sample and Aggregated Embeddings) — Inductive setting

---

Table of Contents

1. Dataset
2. Data Shapes and Sizes
3. Preprocessing
4. Graph Construction
5. Model Architectures
6. Training
7. Results
8. Repository Structure
9. Requirements

---

Dataset

DBpedia-14 Ontology Classification Dataset

The DBpedia dataset is derived from Wikipedia articles and organized into 14 non-overlapping ontology classes. Each sample consists of a title and an abstract (content) extracted from a Wikipedia article.

#	Class Label
0	Company
1	EducationalInstitution
2	Artist
3	Athlete
4	OfficeHolder
5	MeanOfTransportation
6	Building
7	NaturalPlace
8	Village
9	Animal
10	Plant
11	Album
12	Film
13	WrittenWork

• Total Classes: 14
• Source: DBpedia / Hugging Face datasets library (dbpedia_14)
• Text Fields Used: title + content (concatenated)

---

Data Shapes and Sizes

Transductive Setting (GCN)

Split	Samples per Class	Total Samples
Train	2,000	28,000
Test	1,000	14,000
Total Nodes	—	42,000

• TF-IDF Feature Matrix: (42,000 nodes, 5,000 features)
• Node Feature Dimension: 5,000
• Graph Edges (k-NN, K=10): ~438,000 edges
• Graph Type: Homogeneous k-NN cosine similarity graph

Inductive Setting (GraphSAGE)

Split	Samples per Class	Total Samples
Train	2,000	28,000
Test	1,000	14,000
Total Document Nodes	—	42,000

• Vocabulary (Word Nodes): ~5,000 word nodes
• Total Graph Nodes: ~47,000 (42,000 docs + 5,000 words)
• Total Graph Edges: ~1,670,000
  • Doc→Word edges: TF-IDF weighted
  • Word→Word edges: PMI (Pointwise Mutual Information) weighted
• Node Feature Dimension: 5,000 (TF-IDF)

---

Preprocessing

Both pipelines share the following preprocessing steps:

1. Data Loading & Stratified Sampling

• Dataset loaded via Hugging Face datasets (dbpedia_14)
• Stratified sampling applied to ensure class balance:
  • 2,000 samples/class for training
  • 1,000 samples/class for testing

2. Text Cleaning

• Concatenate title and content fields
• Lowercase all text
• Remove punctuation and special characters
• Tokenize using NLTK
• Remove English stopwords (NLTK stopwords corpus)
• Retain only alphabetic tokens

3. TF-IDF Vectorization

• Applied TfidfVectorizer with max_features=5000
• Produces a sparse feature matrix of shape (N_docs, 5000)
• Used as node feature matrix for GNN input

---

Graph Construction

Transductive Graph (GCN)

• Algorithm: k-Nearest Neighbors (k-NN) on TF-IDF cosine similarity
• K: 10 neighbors per node
• Edges: Undirected cosine similarity edges
• Total Edges: ~438,000
• Node Count: 42,000 (train + test combined into one graph)
• All nodes are present during training (transductive)

Inductive / Heterogeneous Graph (GraphSAGE)

• Type: Heterogeneous bipartite-style text graph
• Node Types:
  • Document nodes: 42,000
  • Word nodes: ~5,000 (top TF-IDF vocabulary terms)
• Edge Types:
  • Doc → Word: weighted by TF-IDF score
  • Word → Word: weighted by Pointwise Mutual Information (PMI) over a sliding window
• Total Edges: ~1,670,000
• Test nodes are unseen during training (inductive)

---

Model Architectures

GCN — Transductive

Input: Node features (N, 5000)
  └─> GCNConv Layer 1: 5000 → 256, ReLU, Dropout(0.5)
  └─> GCNConv Layer 2: 256 → 14 (num_classes)
  └─> Log Softmax → Class Prediction

• Framework: PyTorch Geometric
• Optimizer: Adam
• Loss: Negative Log-Likelihood (NLL) / CrossEntropy
• Dropout: 0.5

GraphSAGE — Inductive

Input: Node features (N, 5000)
  └─> SAGEConv Layer 1: 5000 → 256, ReLU, Dropout(0.5)
  └─> SAGEConv Layer 2: 256 → 14 (num_classes)
  └─> Log Softmax → Class Prediction

• Framework: PyTorch Geometric
• Sampling: NeighborLoader for mini-batch neighbor sampling
• Aggregation: Mean aggregation (GraphSAGE default)
• Optimizer: Adam
• Loss: CrossEntropy
• Dropout: 0.5

---

Training

Transductive (GCN)

• Mode: Full-batch (entire graph in memory)
• Train Mask: 28,000 nodes
• Val/Test Mask: 14,000 nodes
• Epochs: 30
• Hardware: GPU (CUDA if available, else CPU)

Inductive (GraphSAGE)

• Mode: Mini-batch with NeighborLoader
• Batch Size: Configurable per run
• Neighbor Sampling: 2-hop neighborhood
• Train Nodes: 28,000 document nodes
• Test Nodes: 14,000 unseen document nodes
• Epochs: 30
• Hardware: GPU (CUDA if available, else CPU)

---

Results

GCN — Transductive

Metric	Value
Best Validation Accuracy	94.93%
Best Epoch	26 / 30
Node Features	TF-IDF (5000-dim)
Graph Edges	~438,000 (k-NN, K=10)

GraphSAGE — Inductive

Metric	Value
Setting	Inductive (unseen test nodes)
Node Features	TF-IDF (5000-dim)
Graph Edges	~1,670,000 (TF-IDF + PMI)
Architecture	2-layer SAGEConv (256 hidden)

Both models demonstrate strong classification performance on the 14-class DBpedia ontology dataset, showcasing the effectiveness of graph-based text representations over traditional flat classifiers.

---

Repository Structure

DbPedia-Enitity-Classification-using-GCNs-and-GraphSage/
├── gcn/
│   └── transductive.ipynb      # GCN transductive classification pipeline
├── GraphSage/
│   └── inductive.ipynb         # GraphSAGE inductive classification pipeline
├── README.md
└── LICENSE

---

Requirements

torch
torch-geometric
transformers
datasets
scikit-learn
nltk
numpy
pandas
matplotlib

Install PyTorch Geometric following the official guide based on your CUDA version.

---

License

MIT License — see LICENSE for details.