CS 848 - Deduplication in RDF

The purpose of the project is to demonstrate the efficacy of graph embeddings for detecting similar nodes in an RDF graph. The code here is made available so that our experiments could be easily replicated.

Datasets

The datasets that we use are available at:

• ground truth
• dblp RDF graph

Project structure

The code is divided into folders by purpose.

• Data - contains code for generating the intermediate datasets from the original data in a form that can be consumed by the rest of the code.
• Baseline - contains the code for running the baseline.
• Graph Embeddings - contains the code for creating graph embeddings.
• Classification - contains the code for creating the classifier model and getting classification results for node pairs.

Running the experiments

Creating datasets.

• Create a smaller RDF dataset from the large dblp_RDF_GRAPH such that only the neighborhood information of the ground truth nodes are available. Run data\crawler.ipynb
  Input: ground truth and dblp RDF graph files.
  Output: dblp_dataset_*.nt file containing the smaller RDF dataset. dblp_*.json file containing a dictionary of ground truth data that was used in creating the smaller dataset.

• Create a list of node pairs along with labelling information. Run data\create_edgelist.py
  Input dblp_*.json file from the previous step.
  Output dblp_*.edges, dblp_*.authors, dblp_*.papers files. dblp_*.edges contains the labelled data along with node ids. The other files map node ids to nodes in the RDF graph.

Running the baseline

Run baseline\run_baseline.ipynb
Input dblp_dataset_*.nt file and dblp_*.edges denoting labelling information.
Output: Accuracy information of the baseline.

Generating graph embeddings

Refer the instructions at graph embedding generation

Classification

Refer the instructions at classification