Text-Attributed-Graph on Patent Data

This repository contains code to build a text-attributed graph from patent data. The graph represents patents as nodes and connects them based on semantic similarity of their textual content.

Features

• Load and preprocess patent data from CSV files
• Generate semantic embeddings using SentenceTransformers
• Build a similarity graph based on customizable thresholds

Patent Graph Dataset

• is located in CS594PatentProject/output folder

Generate Text Embeddings

• Go to GNN/embedding.py
• Set the model you want to use (e.g., all-MiniLM-L6-v2)
• Embeddings will be saved to: output/embeddings/MiniLML6.npy

Data Loader

• Go to GNN/data_loader.py
• Set the appropriate embedding file:

embedding_path = os.path.join(root, "output", "embeddings", "MiniLML6.npy")

Train the GNN Model

To train a GNN on the patent graph

• Open GNN/train.py
• Choose the GNN model by modifying the line: model = GCN(patent_data.num_node_features, 64, len(torch.unique(patent_data.y))) Replace GCN with the specific GNN model you want to use, such as GCN_deep, GraphSAGE, GAT, GIN, etc.

Generate explanation

• Open TAPE_modified and run explanation.py for explanation generation
• Put your huggingface access token in the line: hf_token = "hf_token"

Authors

Homaira Huda Shomee (hshome2@uic.edu) Ataher Sams (asams3@uic.edu)