Deep Learning on Academic Knowledge Graphs

Predicting new facts in a novel semantic graph built on top of the Politecnico di Torino scholarly data

Implementation of an R-GCN based model for node embedding using Deep Graph Library.

data folder contains in XML format the RDF description of the knowledge graphs used.

Relational Graph Convolutional Network (R-GCN)

R-GCN is a kind of GraphConvNet that operates on knowledge graphs.

The main difference with respect to classical GCNs that operates on graphs is that R-GCN operates on multigraphs with labeled edges.

How to run

On a Workstation with:

• 32GB of system memory
• 8GB of video memory

Link predictor training

First train the link predictor.

• Training on GPU and evaluation on CPU:

python3 -u rgcn-linkpredict.py --job="train" --gpu=0 --num-threads=8 --graph-perc=1.0 --train-perc=0.9 --valid-perc=0.05 --test-perc=0.05 --eval-batch-size=80 --graph-batch-size=20000 --n-epochs=6000 --lr=0.001 --regularization=0.5 --evaluate-every=100 --rdf-graph-path="../data/anni2013-2017_no_img_7topics.xml" --load-dataset="input/pkg_dataset.pth" 2>&1 | tee output/output.log

New links evaluation

Use the trained predictor to export candidate links to be added to the graph.

• Evaluation of candidate links to be used for the recommendation system:

python3 -u rgcn-linkpredict.py --job="eval" --gpu=-1 --num-threads=8 --graph-perc="1.0" --eval-batch-size="30" --graph-batch-size="10000" --load-model-state="model_state.pth" --load-data="input/pkg_dataset.pth" --num-scored-triples="30" 2>&1 | tee output/output.log